US 20040024750 A1 Abstract A control system for optimizing a shock absorber having a non-linear kinetic characteristic is described. The control system uses a fitness (performance) function that is based on the physical laws of minimum entropy and biologically inspired constraints relating to mechanical constraints and/or rider comfort, driveability, etc. In one embodiment, a genetic analyzer is used in an off-line mode to develop a teaching signal. The teaching signal can be approximated online by a fuzzy controller that operates using knowledge from a knowledge base. A learning system is used to create the knowledge base for use by the online fuzzy controller. In one embodiment, the learning system uses a quantum search algorithm to search a number of solution spaces to obtain information for the knowledge base. The online fuzzy controller is used to program a linear controller.
Claims(50) 1. A quantum search system for global optimization of a knowledge base and a robust fuzzy control algorithm design for an intelligent mechatronic control suspension system based on quantum soft computing, comprising:
a quantum genetic search module configured to develop a teaching signal for a fuzzy-logic suspension controller, said teaching signal configured to provides a desired set of control qualities over different types of roads; a genetic analyzer module configured to produce a plurality of solutions, at least one solution for each of said different types of roads; and a quantum search module configured to search said plurality of solutions for information to construct said teaching signal. 2. The quantum search system of 3. The quantum search system of 4. The quantum search system of 5. The quantum search system of 6. The quantum search system of 7. The quantum search system of 8. The quantum search system of 9. A control system for a plant comprising:
a neural network configured to control a fuzzy controller, said fuzzy controller configured to control linear controller that controls said plant; a genetic analyzer configured to train said neural network, said genetic analyzer comprising a fitness function that reduces sensor information while reducing entropy production based on biologically-inspired constraints. 10. The control system of 11. The control system of 12. The control system of 13. The control system of 14. The control system of 15. The control system of 16. The control system of 17. The control system of 18. The control system of 19. The control system of 20. The control system of 21. A method for controlling a nonlinear plant by obtaining an entropy production difference between a time derivative dS_{u}/dt of an entropy of the plant and a time derivative dS_{c}/dt of an entropy provided to the plant from a controller; using a genetic algorithm that uses the entropy production difference as a performance function to evolve a control rule in an off-line controller; filtering control rules from an off-line controller to reduce information content and providing filtered control rules to an online controller to control the plant. 22. The method of 23. The method of 24. A self-organizing control system, comprising: a simulator configured to use a thermodynamic model of a nonlinear equation of motion for a plant, a fitness function module that calculates a fitness function based on an entropy production difference between a time differentiation of an entropy of said plant dS_{u}/dt and a time differentiation dS_{c}/dt of an entropy provided to the plant by a linear controller that controls the plant; a genetic analyzer that uses said fitness function to provide a plurality of teaching signals, each teaching signal corresponding to a solution space; a quantum search algorithm module configured to find a global teaching signal from said plurality of teaching signals; a fuzzy logic classifier that determines one or more fuzzy rules by using a learning process and said global teaching signal; and a fuzzy logic controller that uses said fuzzy rules to set a control variable of the linear controller. 25. The self-organizing control system of 26. A control system comprising: a genetic algorithm that provides a plurality of teaching signals corresponding to a plurality of spaces using a fitness function that provides a measure of control quality based on reducing production entropy in each space; a local entropy feedback loop that provides control by relating stability of a plant and controllability of the plant; and a quantum search module to provide a global control teaching signal from said plurality of teaching signals. 27. The control system of 28. The control system of 29. The control system of 30. The control system of 31. An optimization control method for a shock absorber comprising the steps of:
obtaining a difference between a time differential of entropy inside a shock absorber and a time differential of entropy given to said shock absorber from a control unit that controls said shock absorber; and optimizing at least one control parameter of said control unit by using a genetic algorithm and a quantum search algorithm, said genetic algorithm using said difference as a fitness function, said fitness function constrained by at least one biologically-inspired constraint. 32. The optimization control method of 33. The optimization control method of 34. The optimization control method of 35. The optimization control method of 36. The optimization control method of 37. A method for control of a plant comprising the steps of: calculating a first entropy production rate corresponding to an entropy production rate of a control signal provided to a model of said plant; calculating a second entropy production rate corresponding to an entropy production rate of said model of said plant; determining a fitness function for a genetic optimizer using said first entropy production rate and said second entropy production rate; providing said fitness function to said genetic optimizer; providing a teaching output from said genetic optimizer to a quantum search algorithm followed by an information filter; providing a compressed teaching signal from said information filter to a fuzzy neural network, said fuzzy neural network configured to produce a knowledge base; providing said knowledge base to a fuzzy controller, said fuzzy controller using an error signal and said knowledge base to produce a coefficient gain schedule; and providing said coefficient gain schedule to a linear controller. 38. The method of 39. The method of 40. The method of 41. The method of 42. The method of 43. The method of 44. The method of 45. The method of 46. The method of 47. The method of 48. The method of 49. The method of 50. A control apparatus comprising: off-line optimization means for determining a control parameter from an entropy production rate to produce a knowledge base from a compressed teaching signal found by a quantum search algorithm; and online control means for using said knowledge base to develop a control parameter to control a plant.Description [0001] 1. Field of the Invention [0002] The disclosed invention is relates generally to control systems, and more particularly to electronically controlled suspension systems. [0003] 2. Description of the Related Art [0004] Feedback control systems are widely used to maintain the output of a dynamic system at a desired value in spite of external disturbances that would displace it from the desired value. For example, a household space-heating furnace, controlled by a thermostat, is an example of a feedback control system. The thermostat continuously measures the air temperature inside the house, and when the temperature falls below a desired minimum temperature the thermostat turns the furnace on. When the interior temperature reaches the desired minimum temperature, the thermostat turns the furnace off. The thermostat-furnace system maintains the household temperature at a substantially constant value in spite of external disturbances such as a drop in the outside temperature. Similar types of feedback controls are used in many applications. [0005] A central component in a feedback control system is a controlled object, a machine, or a process that can be defined as a “plant”, having an output variable or performance characteristic to be controlled. In the above example, the “plant” is the house, the output variable is the interior air temperature in the house and the disturbance is the flow of heat (dispersion) through the walls of the house. The plant is controlled by a control system. In the above example, the control system is the thermostat in combination with the furnace. The thermostat-furnace system uses simple on-off feedback control system to maintain the temperature of the house. In many control environments, such as motor shaft position or motor speed control systems, simple on-off feedback control is insufficient. More advanced control systems rely on combinations of proportional feedback control, integral feedback control, and derivative feedback control. A feedback control based on a sum of proportional feedback, plus integral feedback, plus derivative feedback, is often referred as PID control. [0006] A PID control system is a linear control system that is based on a dynamic model of the plant. In classical control systems, a linear dynamic model is obtained in the form of dynamic equations, usually ordinary differential equations. The plant is assumed to be relatively linear, time invariant, and stable. However, many real-world plants are time-varying, non-linear, and unstable. For example, the dynamic model may contain parameters (e.g., masses, inductance, aerodynamics coefficients, etc.), which are either only approximately known or depend on a changing environment. If the parameter variation is small and the dynamic model is stable, then the PID controller may be satisfactory. However, if the parameter variation is large or if the dynamic model is unstable, then it is common to add adaptive or intelligent (AI) control functions to the PID control system. [0007] AI control systems use an optimizer, typically a non-linear optimizer, to program the operation of the PID controller and thereby improve the overall operation of the control system. [0008] Classical advanced control theory is based on the assumption that near of equilibrium points all controlled “plants” can be approximated as linear systems. Unfortunately, this assumption is rarely true in the real world. Most plants are highly nonlinear, and often do not have simple control algorithms. In order to meet these needs for a nonlinear control, systems have been developed that use soft computing concepts such as genetic algorithms, fuzzy neural networks, fuzzy controllers and the like. By these techniques, the control system evolves (changes) over time to adapt itself to changes that may occur in the controlled “plant” and/or in the operating environment. [0009] When a genetic analyzer is used to develop a teaching signal for a fuzzy neural network, the teaching signal typically contains unnecessary stochastic noise, making it difficult to later develop an approximation to the teaching signal. Further, a teaching signal developed for one operational condition (e.g. one type of road) may produce poor control quality when used in a different environment (e.g., on a different type of road). [0010] The present invention solves these and other problems by providing a quantum algorithm approach for global optimization of a knowledge base (KB) and a robust fuzzy control algorithm design for intelligent mechatronic control suspension system based on quantum soft computing. In one embodiment, a quantum genetic search algorithm is used to develop a universal teaching signal that provided good control qualities over different types of roads. In one embodiment, a genetic analyzer produces a training signal (solutions) for each type of road, and a quantum search algorithm searches the training signals for information needed to construct the universal training signal. In one embodiment, an intelligent suspension control system, with quantum-logic feedback for the simulation of robust look-up tables is provided. The principle of minimal entropy production rate is used to guarantee conditions for robustness of fuzzy control. Gate design for dynamic simulation of genetic and quantum algorithms is provided. Dynamic analysis and information analysis of the quantum gates leads to “good” solutions with the desired accuracy and reliability. [0011] In one embodiment, the control system uses a fitness (performance) function that is based on the physical laws of minimum entropy and biologically inspired constraints relating to rider comfort, driveability, etc. In one embodiment, a genetic analyzer is used in an off-line mode to develop a teaching signal for one or more roads having different statistical characteristics. Each teaching signal is optimized by the genetic algorithm for a particular type of road. A quantum algorithm is used to develop a single universal teaching signal from the teaching signals produced by the genetic algorithm. An information filter is used to filter the teaching signal to produce a compressed teaching signal. The compressed teaching signal can be approximated online by a fuzzy controller that operates using knowledge from a knowledge base. The control system can be used to control complex plants described by nonlinear, unstable, dissipative models. The control system is configured to use smart simulation techniques for controlling the shock absorber (plant). [0012] In one embodiment, the control system comprises a learning system, such as a neural network that is trained by a genetic analyzer. The genetic analyzer uses a fitness function that maximizes sensor information while minimizing entropy production based on biologically-inspired constraints. [0013] In one embodiment, a suspension control system uses a difference between the time differential (derivative) of entropy from the learning control unit (that is, the entropy production rate of the control signal) and the time differential of the entropy inside the controlled process (or a model of the controlled process, that is, the entropy production rate of the controlled process) as a measure of control performance. In one embodiment, the entropy calculation is based on a thermodynamic model of an equation of motion for a controlled process plant that is treated as an open dynamic system. [0014] The control system is trained by a genetic analyzer that generates a teaching signal for each solution space. The optimized control system provides an optimum control signal based on data obtained from one or more sensors. For example, in a suspension system, a plurality of angle and position sensors can be used. In an off-line learning mode (e.g., in the laboratory, factory, service center, etc.), fuzzy rules are evolved using a kinetic model (or simulation) of the vehicle and its suspension system. Data from the kinetic model is provided to an entropy calculator that calculates input and output entropy production of the model. The input and output entropy productions are provided to a fitness function calculator that calculates a fitness function as a difference in entropy production rates for the genetic analyzer constrained by one or more constraints obtained from rider preferences. The genetic analyzer uses the fitness function to develop set training signals for the off-line control system, each training signal corresponding to an operational environment. A quantum search algorithm is used to reduce the complexity of the teaching signal data across several solution spaces by developing a universal teaching signal. Control parameters (in the form of a knowledge base) from the off-line control system are then provided to an online control system in the vehicle that, using information from the knowledge base, develops a control strategy. [0015] In one embodiment, the invention includes a method for controlling a nonlinear object (a plant) by obtaining an entropy production difference between a time differentiation (dS [0016] In one embodiment, the control method also includes evolving a control rule relative to a variable of the controller by means of a genetic algorithm. The genetic algorithm uses a fitness function based on a difference between a time differentiation of the entropy of the plant (dS [0017] In one embodiment, the invention comprises a self-organizing control system adapted to control a nonlinear plant. The AI control system includes a simulator configured to use a thermodynamic model of a nonlinear equation of motion for the plant. The thermodynamic model is based on an interaction with a Lyapunov function (V), and the simulator uses the function V to analyze control for a state stability of the plant. The control system calculates an entropy production difference between a time differentiation of the entropy of said plant (dS [0018] In yet another embodiment, the invention comprises a new physical measure of control quality based on minimum production entropy and using this measure for a fitness function of genetic algorithm in optimal control system design. This method provides a local entropy feedback loop in the control system. The entropy feedback loop provides for optimal control structure design by relating stability of the plant (using a Lyapunov function) and controllability of the plant (based on production entropy of the control system). The control system is applicable to a wide variety of control systems, including, for example, control systems for mechanical systems, bio-mechanical systems, robotics, electromechanical systems, etc. [0019] In one embodiment, a Quantum Associative Memory (QuAM) with exponential storage capacity is provided. It employs simple spin-1/2 (two-state) quantum systems and represents patterns as quantum operators. In one embodiment, the QuAM is used in a quantum neural network. In one embodiment, a quantum computational learning algorithm that takes advantages of the unique capabilities of quantum computation to produce a neural networks. [0020] The above and other aspects, features, and advantages of the present invention will be more apparent from the following description thereof presented in connection with the following drawings. [0021]FIG. 1 illustrates a general structure of a self-organizing intelligent control system based on soft computing. [0022]FIG. 2 illustrates the structure of a self-organizing intelligent suspension control system with physical and biological measures of control quality based on soft computing [0023]FIG. 3 illustrates the process of constructing the Knowledge Base (KB) for the Fuzzy Controller (FC). [0024]FIG. 4 shows twelve typical road profiles. [0025]FIG. 5 shows a normalized auto-correlation function for different velocities of motion along the road number 9 from FIG. 4. [0026]FIG. 6A is a plot showing results of stochastic simulations based on a one-dimensional Gaussian probability density function. [0027]FIG. 6B is a plot showing results of stochastic simulations based on a one-dimensional uniform probability density function. [0028]FIG. 6C is a plot showing results of stochastic simulations based on a one-dimensional Reileigh probability density function. [0029]FIG. 6D is a plot showing results of stochastic simulations based on a two-dimensional Gaussian probability density function. [0030]FIG. 6E is a plot showing results of stochastic simulations based on a two-dimensional uniform probability density function. [0031]FIG. 6F is a plot showing results of stochastic simulations based on a two-dimensional hyperbolic probability density function. [0032]FIG. 7 illustrates a full car model. [0033]FIG. 8 shows a control damper layout for a suspension-controlled vehicle having adjustable dampers. [0034]FIG. 9 shows damper force characteristics for the adjustable dampers illustrated in FIG. 8. [0035]FIG. 10 shows the structure of an SSCQ from FIG. 2 for use in connection with a simulation model of the full car and suspension system. [0036]FIG. 11 is a flowchart showing operation of the SSCQ. [0037]FIG. 12 shows time intervals associated with the operating mode of the SSCQ. [0038]FIG. 13 is a flowchart showing operation of the SSCQ in connection with the GA. [0039]FIG. 14 shows the genetic analyzer process and the operations of reproduction, crossover, and mutation. [0040]FIG. 15 shows results of variables for the fuzzy neural network. [0041]FIG. 16A shows control of a four-wheeled vehicle using two controllers. [0042]FIG. 16B shows control of a four-wheeled vehicle using a single controller to control all four wheels. [0043]FIG. 17 shows phase plots of β versus dβ/dt for the dynamic and thermodynamic response of the suspension system to three different roads. [0044]FIG. 18 shows phase plots of S versus dS/dt corresponding to the plots in FIG. 17. [0045]FIG. 19 shows three typical road signals, one signal corresponding to a road generated from stochastic simulations and two signals corresponding to roads in Japan. [0046]FIG. 20 shows the general structure of the intelligent control system based on quantum soft computing. [0047]FIG. 21 shows the structure of a self-organizing intelligent control system with physical and biological measures of control quality based on quantum soft computing [0048]FIG. 22 shows inversion about an average. [0049]FIG. 23 shows inversion about average operation as applied to a superposition where all but one of the components are initially identical and of magnitude O(1/{square root}{square root over (N)}) and where one component is initially negative [0050]FIG. 24 shows amplitude distributions resulting from the various quantum gates involved in Grover's quantum search algorithm for the case of three qubits, where the quantum states which are prepared by these gates are (a) |s =|000, (b) H^{(2m)}|s, (c) I_{x} _{ 0 }H^{(2m)}|s, (d)H ^{2m)}I_{x} _{ 0 }H^{(2m)}|s, (e) —I, H^{(2m)}I_{x} _{ 0 }H^{(2m)}|s), (f) —H^{(2m)}I_{s}H^{(2m)}I_{x} _{ 0 }H^{(2m)}|sFIG. 25 shows a comparison of GA and QSA structures.
[0051]FIG. 26 shows the structure of the Quantum Genetic Search Algorithm. [0052]FIG. 27 shows the generalized QGSA with counting of good solutions in look-up tables of fuzzy controllers. [0053]FIG. 28 shows how a quantum mechanical circuit inverts the amplitudes of those states for which the function j(x) is 1. [0054]FIG. 29 shows how the operator Q=-I, U [0055]FIG. 30 is a schematic representation of the quantum oracle U [0056]FIG. 31 shows a quantum mechanical version of the classical-XOR gate as an example for a quantum gate (CNOT gate), where the input state |x, y is mapped into the output state |x, x⊕y.[0057]FIG. 32 shows a variation of coefficients under the (R [0058]FIG. 33 shows fragments of lookup tables generated from different road results. [0059]FIG. 34 shows a general iteration algorithm for information analysis of Grover's algorithm. [0060]FIG. 35 shows a first iteration of the algorithm shown in FIG. 34. [0061]FIG. 36 shows a second iteration of the algorithm shown in FIG. 34. [0062]FIG. 37 shows a scheme Diagram of the QA. [0063]FIG. 38 shows the structure of a Quantum Gate. [0064]FIG. 39 shows methods in Quantum Algorithm Gate Design. [0065]FIG. 40 shows the gate approach for simulation of quantum algorithms using classical computers. [0066]FIG. 41A shows a vector superposition used in a first step of Grover's algorithm. [0067]FIG. 41B shows the superposition from FIG. 41A after applying the operator [0068]FIG. 41C shows the superposition from FIG. 41B after applying the entanglement operator U [0069]FIG. 41D shows the superposition from FIG. 41C after the application of D [0070]FIG. 41E shows the superposition from FIG. 41D after further application of the U [0071]FIG. 41F shows the superposition from FIG. 41E after applying D [0072]FIG. 42 shows Grover's quantum algorithm simulation (Circuit representation and corresponding gate design). [0073]FIG. 43 shows preparation of entanglement operators: a) and b) single solution search; c) for two solutions search; d) for three solutions search. [0074]FIG. 44 shows a quantum gate assembly. [0075]FIG. 45 shows the first iteration of Grover's algorithm execution. [0076]FIG. 46 shows results of the Grover's algorithm execution. [0077]FIG. 47 shows interpretation of Grover' quantum algorithm. [0078]FIG. 48 shows examples of result interpretation of Grover's quantum algorithm. [0079]FIG. 49 shows the circuit for Grover's algorithm where: C is the computational register and M is the memory register; U _{C}. FIG. 50 shows the dependence of the mutual information between the M and the C registers as a function of the number of times.
[0080]FIG. 51 [0081]FIG. 51 [0082]FIG. 52 shows dependence of the required memory for number of qubit. [0083]FIG. 53 shows the time required for a fixed number of iterations for a number of qubit for various Intel Pentium III processors. [0084]FIG. 54 shows the time required for 100 iterations with different internal frequency using an Intel Pentium III CPU. [0085]FIG. 55 shows the time required for fixed number of iterations regarding to number of qubit for Intel Pentium III processors of different internal frequency. [0086]FIG. 56 shows the time required for 10 iterations with different internal frequency of Intel Pentium III processor. [0087]FIG. 57 shows the time required for making one iteration with 11 qubit on PC with 512 MB physical memory. [0088]FIG. 58 shows CPU time required for making one iteration versus the number of qubits. [0089]FIG. 59 shows a dynamic iteration process of a fast quantum search algorithm. [0090]FIG. 60 [0091]FIG. 60 [0092]FIG. 61 shows the structure of a new quantum oracle algorithm in four-dimensional Hilbert space. [0093]FIGS. 62 [0094]FIG. 63 shows general representation of a particular database function f operating on spins I [0095]FIG. 64 shows quantum search algorithm in spin Liouville space. [0096]FIG. 65 shows general representation of a particular database function f operating on spins I [0097]FIG. 67 shows effects of D operation: (a) States before operation; (b) States after operation. [0098]FIG. 68 shows finding 1 out of N items. (a) Uniform superposition is prepared initially. Every item has equal amplitude (1/{square root}{square root over (N)}); (b) Oracle U [0099]FIG. 69 shows geometric interpretation of the iterative procedure. [0100]FIG. 70 shows the design process of KB for fuzzy P-controller with QGSA. [0101]FIG. 71 shows a quantum genetic search algorithm structure. [0102]FIG. 72 shows a geometrical interpretation of a new quantum oracle. [0103]FIG. 73 shows a gate structure of a new quantum oracle. [0104]FIG. 74 shows a gate structure of quantum genetic search algorithm. [0105] In the drawings, the first digit of any three-digit element reference number generally indicates the number of the figure in which the referenced element first appears. The first two digits of any four-digit element reference number generally indicate the figure in which the referenced element first appears. [0106]FIG. 1 is a block diagram of a control system [0107] An output of the entropy-calculation module [0108] The GA [0109] Using a set of inputs, and the fitness function [0110] Chromosomes that are more fit are kept (survive) and chromosomes that are less fit are discarded (die). New chromosomes are created to replace the discarded chromosomes. The new chromosomes are created by crossing pieces of existing chromosomes and by introducing mutations. [0111] The PID controller [0112] Evaluating the motion characteristics of a nonlinear plant is often difficult, in part due to the lack of a general analysis method. Conventionally, when controlling a plant with nonlinear motion characteristics, it is common to find certain equilibrium points of the plant and the motion characteristics of the plant are linearized in a vicinity near an equilibrium point. Control is then based on evaluating the pseudo (linearized) motion characteristics near the equilibrium point. This technique is scarcely, if at all, effective for plants described by models that are unstable or dissipative. [0113] Computation of optimal control based on soft computing includes the GA [0114] In order to realize an intelligent mechatronic suspension control system, the structure depicted on FIG. 1 is modified, as shown on FIG. 2 to produce a system [0115] As shown in FIG. 3, realization of the structure depicted in FIG. 2 is divided into four development stages. The development stages include a teaching signal acquisition stage [0116] The teaching signal acquisition stage [0117] The output of the stage [0118] Behavior of the control system is obtained from the output of the GA [0119] The third stage [0120] The output of the third stage [0121] In the fourth stage [0122] To summarize, the development of the KB for an intelligent control suspension system includes: [0123] I. Obtaining a stochastic model of the road or roads. [0124] II. Obtaining a realistic model of a car and its suspension system. [0125] III. Development of a Simulation System of Control Quality with the car model for genetic algorithm fitness function calculation, and introduction of human needs in the fitness function. [0126] IV. Development of the information compressor (information filter). [0127] V. Approximation of the teaching signal with a fuzzy logic classifier system (FLCS) and obtaining the KB for the FC [0128] VI. Verification of the KB in experiment and/or in simulations of the full car model with fuzzy control [0129] I. Obtaining Stochastic Models of the Roads [0130] It is convenient to consider different types of roads as stochastic processes with different auto-correlation functions and probability density functions. FIG. 4 shows twelve typical road profiles. Each profile shows distance along the road (on the x-axis), and altitude of the road (on the y-axis) with respect to a reference altitude. FIG. 5 shows a normalized auto-correlation function for different velocities of motion along the road number 503 shows the normalized auto-correlation function for =5 meter/sec, and a curve 504 shows the normalized auto-correlation function for =10 meter/sec.
[0131] The results of statistical analysis of actual roads, as shown in FIG. 4, show that it is useful to consider the road signals as stochastic processes using the following three typical auto-correlation functions. _{1} τ; (1.2) [0132] where α [0133] α [0134] For convenience, the roads are divided into three classes:
[0135] The presented auto-correlation functions and its parameters are used for stochastic simulations of different types of roads using forming filters. The methodology of forming filter structure can be described according to the first type of auto-correlation functions (1.1) with different probability density functions. [0136] Consider a stationary stochastic process X(t) defined on the interval [x [0137] Let the spectral density be of the following low-pass type:
[0138] where σ [0139] where α is the same parameter in (2.1), B(t) is a unit Wiener process, and the coefficients −αX and D(X) are known as drift and the diffusion coefficients, respectively. To demonstrate that this is the case, multiply (2.2) by X(t−τ) and take the ensemble average to yield
[0140] where R(τ) is the correlation function of X(t), namely, R(τ)=E [X(t−τ)X(t)]. Equation (2.3) has a solution [0141] in which A is arbitrary. By choosing A=σ [0142] Now it is useful to determine D(X) so that X(t) possesses a given stationary probability density p(x). The Fokker-Planck equation, governing the probability density p(x) of X(t) in the stationary state, is obtained from equation (2.2) as follows:
[0143] where G is known as the probability flow. Since X(t) is defined on [x [0144] Integration of equation (2.6) results in
[0145] where C is an integration constant. To determine the integration constant C, two cases are considered. For the first case, if x [0146] Function D [0147] The Ito type stochastic differential equation (2.2) may be converted to that of the Stratonovich type as follows:
[0148] where ξ(t) is a Gaussian white noise with a unit spectral density. Equation (2.9) is better suited for simulating sample functions. Some illustrative examples are given below. [0149] Assume that X(t) is uniformly distributed, namely
[0150] Substituting (2.10) into (2.8) [0151] In this case, the desired Ito equation is given by [0152] It is of interest to note that a family of stochastic processes can be obtained from the following generalized version of (2.12): [0153] Their appearances are strikingly diverse, yet they share the same spectral density (2.1). [0154] Let X(t) be governed by a Rayleigh distribution [0155] Its centralized version Y(t)=X(t)−2/γ has a probability density [0156] From equation (2.8),
[0157] The Ito equation for Y(t) is
[0158] and the correspondence equation for X(t) in the Stratonovich form is
[0159] Note that the spectral density of X(t) contains a delta function (4/γ [0160] Consider a family of probability densities, which obeys an equation of the form
[0161] Equation (2.19) can be integrated to yield [0162] where C [0163] Several special cases may be noted. Let [0164] where γ can be arbitrary if δ>0. Substitution of equation (2.22) into equation (2.8) leads to
[0165] where erfc(y) is the complementary error function defined as
[0166] The case of γ<0 and δ>0 corresponds to a bimodal distribution, and the case of γ>0 and δ=0 corresponds to a Gaussian distribution. [0167] The Pearson family of probability distributions corresponds to
[0168] In the special case of a [0169] From the results of statistical analysis of forming filters with auto-correlation function (1.1) one can describe typical structure of forming filters as in Table 2.1:
[0170] The structure of a forming filter with an auto-correlation function given by equations (1.2) and (1.3) is derived as follows. A two-dimensional (2D) system is used to generate a narrow-band stochastic process with the spectrum peak located at a nonzero frequency. The following pair of Ito equations describes a large class of 2D systems: [0171] where B [0172] For a system to be stable and to possess a stationary probability density, is required that a [0173] where R [0174] Differential equations (3.2) in the time domain can be transformed (using the Fourier transform) into algebraic equations in the frequency domain as follows
[0175] where {overscore (R)} [0176] Then the spectral density S [0177] where Re denotes the real part. [0178] Since R [0179] and equation (3.3) is obtained using this relation. [0180] Solving equation (3.3) for {overscore (R)} [0181] where A [0182] Expression (3.5) is the general expression for a narrow-band spectral density. The constants a [0183] Forming filters for simulation of non-Gaussian stochastic processes can be derived as follows. The Fokker-Planck-Kolmogorov (FPK) equation for the joint density p(x [0184] If such D [0185] where ξ [0186] Filters (3.1) and (3.6) are non-linear filters for simulation of non-Gaussian random processes. Two typical examples are provided. CL EXAMPLE 1 [0187] Consider two independent uniformly distributed stochastic process x [0188] −Δ [0189] In this case, from the FPK equation, one obtains
[0190] which is satisfied if [0191] D [0192] The two non-linear equations in (3.6) are now
[0193] which generate a uniformly distributed stochastic process x, (t) with a spectral density given by (3.5). [0194] Consider a joint stationary probability density of x [0195] p(x [0196] A large class of probability densities can be fitted in this form. In this case
[0197] The forming filter equations (3.6) for this case can be described as following
[0198] If σ [0199] The stochastic differential equation for the variable x [0200] These equations can be integrated using two different algorithms: Milshtein; and Heun methods. In the Milshtein method, the solution of stochastic differential equation (4.1) is computed by the means of the following recursive relations:
[0201] where η [0202] The second term in equation (4.2) is included because equation (4.2) is interpreted in the Stratonovich sense. The order of numerical error in the Milshtein method is δt. Therefore, small δt (i.e., (δt=1×10 [0203] The Heun method is based on the second-order Runge-Kutta method, and integrates the stochastic equation by using the following recursive equation:
[0204] where [0205] y [0206] The Heun method accepts larger δt than the Milshtein method without a significant increase in computational effort per step. The Heun method is usually used for σ [0207] The time step δt can be chosen by using a stability condition, and so that averaged magnitudes do not depend on δt within statistical errors. For example, δt=5×10 [0208] Table 3.1 summarizes the stochastic simulation of typical road signals.
[0209]FIG. 7 shows a vehicle body [0210] { [0211] { [0212] { [0213] { [0214] { [0215] { [0216] Expressions for the entropy production of the suspension system shown in FIG. 7 are developed in U.S. application Ser. No. 09/176,987 hereby incorporated by reference in its entirety. [0217]FIG. 8 shows the vehicle body [0218]FIG. 9 shows damper force characteristics as damper force versus piston speed characteristics when the rotary valve is placed in a hard damping position and in a soft damping position. The valve is controlled by the stepping motor to be placed between the soft and the hard damping positions to generate intermediate damping force. [0219] The SSCQ [0220] The Timer [0221] Road signal generator 1010 generates a road profile. The road profile can be generated from stochastic simulations as described above in connection with FIGS. [0222] The simulation model [0223] The SSCQ [0224] The following designations regarding time moments are used herein:
[0225]FIG. 11 is a flowchart showing operation of the SSCQ [0226] 1. At the initial moment (T=0) the SSCQ [0227] 2. The simulation model [0228] 3. The output X [0229] 4. The time interval T is incremented by T [0230] 5. The sequence 1-4 is repeated a desired number of times (that is while T<T [0231] Regarding step 1 above, the SSCQ block has two operating modes: [0232] [0233] 2. Extraction of the output CGS [0234] The operating mode of the SSCQ [0235]FIG. 13 is a flowchart [0236] The structure of the output buffer
[0237] The output buffer [0238] Two similar models are used. The simulation model [0239] Numerical integration using methods of type (1) is very precise, but time-consuming. Methods of type (2) are typically faster, but with smaller precision. During each SSCQ call in the GA mode, the GA [0240] The fitness function calculation block [0241] The fitness function [0242] where: [0243] i denotes indexes of state variables which should be minimized by their absolute value; j denotes indexes of state variables whose control error should be minimized; k denotes indexes of state variables whose frequency components should be minimized; and w [0244] Extraction of frequency components can be done using standard digital filtering design techniques for obtaining the filter parameters. Digital filtering can be provided by a standard difference equation applied to elements of the matrix X [0245] where a, b are parameters of the filter, N is the number of the current point, and n [0246] In one embodiment, the GA [0247] The reproduction process biases the search toward producing more fit members in the population and eliminating the less fit ones. Hence, a fitness value is first assigned to each string (chromosome) the population. One simple approach to select members from an initial population to participate in the reproduction is to assign each member a probability of selection on the basis of its fitness value. A new population pool of the same size as the original is then created with a higher average fitness value. [0248] The process of reproduction simply results in more copies of the dominant or fit designs to be present in the population. The crossover process allows for an exchange of design characteristics among members of the population pool with the intent of improving the fitness of the next generation. Crossover is executed by selecting strings of two mating parents, randomly choosing two sites. [0249] Mutation safeguards the genetic search process from a premature loss of valuable genetic material during reproduction and crossover. The process of mutation is simply to choose few members from the population pool according to the probability of mutation and to switch a 0 to 1 or vice versa at randomly sites on the chromosome. [0250]FIG. 14 illustrates the processes of reproduction, crossover and mutation on a set of chromosomes in a genetic analyzer. A population of strings is first transformed into decimal codes and then sent into the physical process [0251] The Fuzzy Logic Control System (FLCS) [0252] As it was described above, the output of the SSCQ is a teaching signal K [0253] Furthermore, if entire blocks of independent messages are coded together, then the mean number {overscore (L)} of bits per message can be brought arbitrary close to H(A). [0254] This noiseless coding theorem shows the importance of the Shannon entropy H(A) for the information theory. It also provides the interpretation of H(A) as a mean number of bits necessary to code the output of A using an ideal code. Each bit has a fixed ‘cost’ (in units of energy or space or money), so that H(A) is a measure of the tangible resources necessary to represent the information produced by A. [0255] In classical statistical mechanics, in fact, the statistical entropy is formally identically to the Shannon entropy. The entropy of a macrostate can be interpreted as the number of bits that would be required to specify the microstate of the system. [0256] Suppose x [0257] This Standard result is known as the weak law of large numbers. A sufficiently long sequence of independent, identically distributed random variables will, with a probability approaching unity, have an average that is close to mean of each variable. [0258] The weak law can be used to derive a relation between Shannon entropy H(A) and the number of ‘likely’ sequences of N identical random variables. Assume that a message source A produces the message a with probability p(a). A sequence α=a [0259] From the weak law, it follows that, if ε, δ>0, then for sufficient large N
[0260] for N sequences of α. It is possible to partition the set of all N sequences into two subsets: [0261] a) A set A of “likely” sequences for which
[0262] b) A set of ‘unlikely’ sequences with total probability less than a, for which this inequality fails. [0263] This provides the possibility to exclude the ‘unlikely’ information from the set A which leaves the set of sequences A, with the same information amount as in set Abut with a smaller number of sequences. [0264] The FNN [0265] An example of a KB of a suspension system fuzzy controller obtained using the FNN [0266] The type of fuzzy inference system in this case is a zero-order Sugeno-Takagi Fuzzy inference system. In this case the rule base has the form presented in the list below. [0267] IF ANT [0268] IF ANT [0269] . . . [0270] IF ANT [0271] In the example above, when there are only 25 possible combinations of input membership functions, so it is possible to use all the possible rules. However, when the number of input variables is large, the phenomenon known as “rule blow” takes place. For example, if number of input variables is 6, and each of them has 5 membership functions, then the total number of rules could be: N=5 [0272] The FC [0273] Fuzzyfication is a transferring of numerical data from sensors into a linguistic plane by assigning membership degree to each membership function. The information of input membership function parameters stored in the knowledge base of fuzzy controller is used. [0274] Fussy inference is a procedure that generates linguistic output from the set of linguistic inputs obtained after fuzzyfication. In order to perform the fuzzy inference, the information of rules and of output membership functions from knowledge base is used. [0275] Defuzzyfication is a process of converting of linguistic information into the digital plane. Usually, the process of defuzzyfication include selecting of center of gravity of a resulted linguistic membership function. [0276] Fuzzy control of a suspension system is aimed at coordinating damping factors of each damper to control parameters of motion of car body. Parameters of motion can include, for example, pitching motion, rolling motion, heave movement, and/or derivatives of these parameters. Fuzzy control in this case can be realized in the different ways, and different number of fuzzy controllers used. For example, in one embodiment shown in FIG. 16A, fuzzy control is implemented using two separate controllers, one controller for the front wheels, and one controller for the rear wheels, as shown in FIG. 16A, where a first fuzzy controller [0277] Quantum Searching [0278] As discussed above, the GA uses a global search algorithm based on the mechanics of natural genetics and natural selection. In the genetic search, each design variable is presented by a finite length binary string and the set of all possible solutions is so encoded into a population of binary strings. Genetic transformations, analogous to biological reproduction and evolution, are subsequently used to vary and improve the encoded solutions. Usually, three main operators, reproduction, crossover and mutation are used in the genetic search. [0279] The reproduction process is one that biases the search toward producing more fit members in the population and eliminating the less fit ones. Hence, a fitness value is first assigned to each string in the population. One simple approach to select members from an initial population to participate in the reproduction is to assign each member a probability of being selected, on the basis of its fitness value. A new population pool of the same size as the original is then created with a higher average fitness value. The process of reproduction results in more copies of the dominant design to be present in the population. [0280] The crossover process allows for an exchange of design characteristics among members of the population pool with the intent of improving the fitness of the next generation. Crossover is executed, for example, by selecting strings of two mating parents, randomly choosing two sites on the strings, and swapping strings of 0's and 1's between these chosen sites. [0281] Mutation helps safeguard the genetic search process from a premature loss of valuable genetic material during reproduction and crossover. The process of mutation involves choosing a few members from the population pool on the basis of their probability of mutation and switch 0 to 1 or vice versa at a randomly selected mutation rate on the selected string. [0282] For 1-point crossover χ and mutation μ, the Walsh-Hadamard transform of the (2-bit representation) mixing matrix is given by:
[0283] The matrix {circumflex over (M)} is sparse, containing nine non-zero entries. The Walsh-Hadamard transform of the twist of the (2-bit representation) mixing matrix is given by:
[0284] The mixing matrix is lower triangular. With the above matrix representation of a GA, it is possible to describe the GA in terms of a quantum gate as described in more detail below. [0285] Typically, the GA uses function evaluations alone and does not require function derivatives. While derivatives contribute to a faster convergence towards an optimum, derivatives may also direct the search towards a local optimum. Furthermore, since the search proceeds from several points in the design space to another set of design points, the GA method has a higher probability of locating a global minimum as opposed to those schemes that proceed from one point to another. In addition, genetic algorithms often work on a coding of design variables rather than variables themselves. This allows for an extension of these algorithms to a design space having a mix of continuous, discrete, and integer variables. These properties and the gate representation of GA are used below in a quantum genetic search algorithm. [0286] As discussed above, FIG. 1 shows an intelligent control suspension system [0287] For soft computing systems based on a genetic algorithm, there is very often no real control law in the classic control sense, but rather, control is based on a physical control law such as minimum entropy production. This allows robust control because the GA, combined with feedback, guarantee robustness. However, robust control is not necessarily optimal control. [0288] For random excitations with different statistical properties the GA attempts to find a global optimum solution for a given solution space. The GA produces look-up tables for the FC [0289] A new solution can be found by repeating the simulation with the GA and finding another single space solution with the entropy-based fitness function for the fuzzy controller with non-Gaussian excitation on the control object. As result, it is possible to generate different look-up tables for the fuzzy controller 143 for different road classes with different types of statistical characteristics. [0290] The control system [0291] The statistical characteristics of the road signals produce different responses in the dynamic suspension system and as a result, require different control solution strategies. [0292]FIGS. 17 and 18 illustrate the dynamic and thermodynamic response of the suspension system (plant) to the above-mentioned excitations. Curves [0293] The system responses from the roads with the same characteristics are similar, which means that the GA [0294] The GA 131 are provided to inputs of the QGSA 2001, and a universal output solution (teaching signal) K_{0 }from the QGSA 2001 is provided to the FNN 142. In FIG. 21, the K_{1 }. . . K_{n }solutions from the GA 131 are provided to inputs of an information compressor 2101 and compressed solutions K_{1 }. . . K_{n }are provided to the QGSA 2001. The information compressor 2101 performs information filtering similar to that provided by the information filter 241.
[0295] The QGSA [0296] Superposition is fundamental in quantum mechanics and when applied to composite quantum systems it leads to the notion of entanglement. Interference on the other hand is usually used for classical mechanics. The superposition, entanglement and interference operators are used as three separate terms because they are standard components of a quantum gate. [0297] A quantum computation involves preparing an initial superposition of states, operating on those states with a series of unitary matrices, and then making a measurement to obtain a definite final answer. The amplitudes of the states determine the probability that this final measurement produced a desired result. Using this as a search method, one can obtain each final state with some probability, and some of these states will be solutions. Thus, this is a probabilistic computation in which at each trial produces some probability of a solution, but no guarantee of a solution. This means the quantum search method is incomplete in that it can find a solution if one exists but can never guarantee a solution in one does not exist. [0298] A useful conceptual view is provided by the path integral approach to quantum mechanics. In this view, the final amplitude of a given state is obtained by summing over all possible paths that produce that state, weighted by suitable amplitudes. In this way, various possibilities involved in a computation can interfere with each other, either constructively or destructively. This differs from the classical combination of probabilities of different ways to reach the same outcome, where the probabilities are simply added, giving no possibility for interference. [0299] Consider, for example, a computation that depends on a single choice. The possible choice can be represented as an input bit with value 1 or −1. Assume that the result of the computation from a choice is also a single value, 1 or −1, representing, for example, some consequence of the choice. If one is interested in whether the two results are the same, classically this requires evaluating each choice separately. With a quantum computation one can instead prepare a superposition of the inputs,
[0300] using the matrix H, then do the evaluation to give
[0301] where f [0302] to obtain
[0303] Now if both choices give the same value for f, this result is ±|0 so the final measurement process will give 0. Conversely, if the values are different, this resulting state is ±|1 and the measurement gives 1. Thus, with the effort required to compute one value classically, it is possible to determine definitely whether the two evaluations are the same or different.[0304] In this example, it was assumed that one could arrange to be in a single state at the end of the computation and hence have no probability for obtaining the wrong answer by the measurement. This result is viewed as summing over the different paths; e.g., the final amplitude for |0 , was the sum over the paths |0→|0→|0 and |0→|1→|0. The various formulations of quantum mechanics, involving operators, matrices or sums over paths are equivalent but suggest different thought processes when constructing possible quantum algorithms.[0305] One example of a robust quantum search algorithm is the algorithm due to Grover. In each iteration of the Grover's quantum search algorithm, there are two steps: 1) a selective inversion of the amplitude of the marked state, which is a phase rotation of π of the marked state; 2) an inversion about the average of the amplitudes of all basis states (both of these operations are described in Appendix 2). The second step can be realized by two Walsh-Hadamard transformations and a rotation of π on all basis states different from |0 .[0306] The success of Grover's quantum search algorithm and its multi-object generalization is attributable to two main sources: 1) the notion of amplitude amplification; and 2) the reduction to invariant sub-spaces of low dimension for the unitary operators involved. Indeed, the second of these can be said to be responsible for the first: A proper geometrical formulation of the process shows that the algorithm operates primarily within a two-dimensional real sub-space of the Hilbert space of quantum states. Since the state vectors are normalized, the state is confined to a one-dimensional unit circle and (if moved at all) initially has nowhere to go except toward the place where the amplitude for the sought-for state is maximized. This accounts for the robustness of Grover's quantum search algorithm—that is, the fact that Grover's original choice of initial state and of the Walsh-Hadamard transformation can be replaced by (almost) any initial state and (almost) any unitary transformation. [0307] In general form, Grover's quantum search algorithm is a series of rotations in an SU(2) space spanned by |x [0308] Each iteration rotates the state vector of the quantum computer system an angle
[0309] towards the |x _{0} basis in the SU(2) space, but the angle of rotation is smaller than ψ. For reasons of efficiency, the phase rotation π is generally used. The inversion of the amplitude of the marked state in step 1 is replaced by a rotation through an angle between 0 and π to produce a smaller angle of SU(2) rotation towards the end of a quantum search calculation so that the amplitude of the marked state in the computer system state vector is exactly 1. When the rotation of the phase of the marked state is not π, one cannot simply construct a quantum search algorithm. In vicinity of π, the Grover's algorithm still works, though the height of the norm cannot reach 1. But it can still reach a relatively large value. This shows that Grover's algorithm is robust with respect of phase rotation to π. Grover's quantum search algorithm has good tolerance for a phase rotating angle near π. In other words, a small deviation from π will not destroy the algorithm. This is useful, as an imperfect gate operation may lead to a phase rotation not exactly equal to π.
[0310] From the mathematical point of view, a large class of problems can be specified as search problems of the form “find some x such that P(x) is true” for some predicate P. Such problems range from sorting to graph coloring to database search, etc. For example: [0311] Given an ne lement vector A, find a permutation π on [1, . . . , n] such that ∀1≦i<n:A [0312] Given a graph (V, E) with n vertices V and e edges E [0313] For certain types of problems, where there is some problem structure that can be exploited, efficient algorithms are known. Many search problems, such as constraint satisfaction problems involving graph colorability, or searching an alphabetized list, have structured search spaces in which full solutions can be built from smaller partial solutions. But in the general case with no structure, randomly testing predicates P(x [0314] Quantum algorithms that use the problem structure in a similar way to classical heuristic search algorithms can be useful. One problem with this approach is that the introduction of problem structure often makes the algorithms complicated enough that it is hard to determine the probability that a single iteration of the algorithm will give a correct answer. Therefore it is difficult to know how efficient structured quantum algorithms are. Classically, the efficiency of heuristic algorithms is estimated by empirically testing the algorithm. But, as there is an exponential slow down when simulating a quantum computer on a classical one, empirical testing of quantum algorithms is currently infeasible except in small cases. [0315] Grover's algorithm searches an unstructured list of size N. Let n be such that 2 U [0316] where “True” is encoded as 1. [0317] The first step is the standard step for quantum computing: Compute P for all possible inputs x [0318] of all 2 [0319] For any x [0320] but since its amplitude is
[0321] the probability that a measurement of the 1 ‘-’ superposition produces x [0322] so as to greatly increase the amplitude of vectors |x, 0) for which the predicate is false. [0323] Once such a transformation of the quantum state has been performed, one can simply measure the last qubit of the quantum state, which represents P(x). Because of the amplitude change, there is a high probability that the result will be 1. If this is the case, the measurement has projected the state
[0324] onto the subspace
[0325] where k is the number of solutions. Further, measurement of the remaining bits will provide one of these solutions. If the measurement of qubit P(x) yields 0, then the whole process is started over and the superposition
[0326] is computed again. [0327] Grover's algorithm includes of the following steps: [0328] 1. Prepare a register containing a superposition of all of the possible values x [0329] 2. Compute P(x [0330] 3. Change the amplitude a [0331] 4. Apply inversion about the average to increase the amplitude of x [0332] 5. Repeat steps 2 through
[0333] times. [0334] 6. Read the result. [0335] Grover's algorithm is optimal up to a constant factor, no quantum algorithm can perform an unstructured search faster. If there is only a single x [0336] iterations of steps 2 through 4 the failure rate, is 0.5. After iterating
[0337] times the failure rate drops to 2 [0338] iterations the failure rate is close to 1. [0339] There are many classical algorithms in which a procedure is repeated over and over again for ever better results. Repeating quantum procedures may improve results for a while, but after a sufficient number of repetitions the results will get worse again. Quantum procedures are unitary transformations, which are rotations of complex space, and thus while a repeated applications of a quantum transform may rotate the state closer and closer to the desired state for a while, eventually it will rotate past the desired state to get farther and farther from the desired state. Thus, to obtain useful results from a repeated application of a quantum transformation, it is useful to know when to stop. [0340] The loop in steps 3-5 above is the heart of the Grover search algorithm. Each iteration of this loop increases the amplitude in the desired state by
[0341] as a result in O({square root}{square root over (N)}) repetitions of the loop, the amplitude and hence the probability of being in the desired state reach O(1). To show that the amplitude increases by
[0342] in each repetition, it is first useful show that the diffusion transform, D, can be interpreted as an inversion about an average. A simple inversion is a phase rotation operation, and it is unitary. The inversion about average operation (as developed Appendix 2) is also a unitary operation and is equivalent to the diffusion transform D as used in steps [0343] Let α denote the average amplitude over all state, i.e., if α [0344] As a result of the operation D, the amplitude in each state increases (decreases) so that after this operation it is as much below (above) α, as it was above (below) α before the operation (see FIG. 23). The diffusion transform D is defined as follows:
[0345] D can be represented in the form D=−I+2P, where operator I is the identity matrix and P is a projection matrix with P [0346] In order to see that D is the inversion about average, consider what happens when D acts on an arbitrary vector {overscore (v)}. Expressing D as −I+2P, it follows that: [0347] By the discussion above, each component of the vector P{right arrow over (v)} is A, where A is the average of all components of the vector {overscore (v)}. Therefore, the i-th component of the vector D{overscore (v)} is given by (−v [0348] Next consider the situation, shown in FIG. 23, when this operator is applied to a vector with each of the components, except one, having an amplitude equal to C/{square root}{square root over (N)} a where C list between ½ and 1. The one component that is different has an amplitude of (−{square root}{square root over (1−C [0349] The quantum search algorithm can also be expressed as follows: Given a function f(x [0350] find a target element by using the least number of calls to the function f(x [0351] Grover's algorithm can be generalized as follows. First, form a Hilbert space with an orthonormal basis element for each input x _{i} . The basis of input eigenstates is called the measurement basis. Let N=|χ| be the cardinality of χ. The function call is to be implemented by a unitary operator that acts as follows:
| |y→|x _{i} |y⊕f(x ^{i}) (8.2)
[0352] where |y is either |0 or |1. By acting on[0353] with this operator construct the state
[0354] where the r measurement basis states |t _{i} are the non-target states. Disregarding the state
[0355] then the phase of the target states has been inverted. Hence, the unitary operator above is equivalent to the operator
[0356] (It is not necessary to know what the target states are a priori.) Next, construct the operator Q defined as
[0357] Where |a can be thought of as the averaging state. Different choices of |a give rise to different unitary operators for performing amplitude amplification. In the original Grover algorithm, the state |a was chosen to be[0358] and was obtained by applying the Walsh-Hadamard operator, U, to a starting eigenstate |s , i.e., |a=U|s. Hence, the operation (2|a a|−1), called inversion about the average, is equivalent to −UI_{s}U^{+} with U being the Walsh-Hadamard operator and I_{s }being 1−2|s s|.
[0359] By knowing more about the structure of the problem one can choose other vectors |a that will allow finding a target state faster.[0360] Fortunately, in order to determine what action the operator Q performs, it is sufficient to focus on a two-dimensional subspace. The basis vectors of this subspace can be written as
[0361] It is observed that |t is the normalized projection of |a onto the space of target states and |a′ is the normalized projection of |a onto the space orthogonal to |t.[0362] The rest of the Hilbert space (i.e., the space orthogonal to |t and |a′) can be broken up into the space of target states (S_{T}) and the space of non-target states (S_{L}). Q can be written as
t|+|a′ a′|)+sinφ(|t−|a′ t|)+I _{T} −I _{L}, φ≡cos^{−1}[1−2v ^{2}] (8.9)
[0363] where I [0364] An arbitrary starting superposition |s for the algorithm can be written as| =α|+βe ^{ib} |a′+|φ _{t} +|φ_{I} (8.10)
[0365] where the states |φ _{l} (which have a norm less than one if the state |s is to be properly normalized overall) are the components of |s in (S_{T}) and (S_{L}) respectively. Also, α, β and b are positive real numbers. After n applications of Q on an arbitrary starting superposition |s one obtains
nφ)+βe ^{ib }sin(nφ)]|t+[βe ^{ib }cos(nφ)+α sin(nφ)]|a′+|φ_{l} +(−1)^{n}|φ_{l} (8.11)
[0366] Measuring this state provides the probability of success (i.e., measuring a target state) as given by two terms. [0367] The first term is the magnitude squared of Q _{T}. This magnitude is (φ_{t}|φ_{t} and is unchanged by Q.
[0368] The value g(n) is the magnitude squared of the coefficient of |t , which is given by[0369] where
[0370] This is the term that is affected by Q and is the term to be maximized. The total probability of success after n iterations of Q acting on |s is +g(n) (8.13)
[0371] Assuming that n is continuous (an assumption that is justified below) the maxima of g(n), and hence the maxima of the probability of success of Grover's algorithm, are given by the following.
[0372] The value of g(n) at these maxima is given by
[0373] In practice, the optimal n must be an integer and typically the n [0374] around n [0375] is the probability of measuring a target state after n [0376] Grover's algorithm provides for searching a single element in an unsorted database (DB). The above description is presented in a way that makes possible the generalization of the algorithm to perform multi-object search in an unstructured DB. [0377] The Grover's quantum search algorithm was developed for searching a single element in an unsorted database containing N>>1 items and treated the following abstract problem: given a Boolean function f(w)=1, w=1, . . . , N, which is known to be zero for all w except at a single point, say at w=a, where f(a)=1; find the value a. The function can be treated as “oracle” or “black box” wherein all that is known about it is its output for any input. On a classical computer it is necessary to evaluate the function
[0378] times on average to find the answer to this problem. In contrast, Grover's quantum search algorithm finds a solution in O({square root}{square root over (N)}) steps. [0379] The quantum-mechanical statement of the above search problem is: given an orthogonal basis |w :w=1, 2, . . . , N; single out the basis element |a for which f(a)=1. Each |w is to be an eigenstate of the qubits making up the quantum computing. If N=2^{n}, then n qubits will be needed. At T=0, prepare the state of the system |ψ in a superposition of the state {|w}, each with the same probability:
[0380] By the Graham-Schmidt construction, extend |a to an orthonormal basis for the sub-space spanned by |a and |s. That is, introduce a normalized vector |r orthogonal to |a,[0381] and find that the initial state has the representation
[0382] Following Grover's quantum search algorithm, now define the unitary operator of inversion about the average, I [0383] The only action of this operator is to flip the sign of the state |s ; that is, I_{s}|s=−|s but I_{s}|v=|v if s|v=0.
[0384] I [0385] or, with respect to the orthonormal basis, the operator (8.18) can be represented by the orthogonal real unitary matrix
[0386] Similarly, define the operator I _{a}|a=−|a. In terms of the oracle function f,
^{f(w)} |w
[0387] for each |w in the original basis for the full state space of the quantum computing.[0388] Therefore, to execute the operation I [0389] A simple “Grover's iteration” is the unitary operator U≅−I [0390] The fact that the matrix element (a|U|s) is nonzero can be used to reinforce the probability amplitude of the unknown state |a . Using U as a unitary search operation, then after m>>1 trials the value a|U^{m}|s can be evaluated as follows:
[0391] Setting | a|U^{m}|s|^{2}=|cos(mθ−α)|^{2}=1, one can maximize the amplitude of U^{m}|s in the state |a; thus
[0392] (if no integer satisfies this equation exactly, take the closest one.) [0393] When N is large,
[0394] and obtain
[0395] Therefore, after m=O({square root}{square root over (N)}) trials, the state |a will be projected out, which is precisely Grover's result. By observing the qubits, a is determined. By constructive interference, it is possible to construct |a. Since m only approximately satisfies (8.19), there is a small chance of getting a “bad” a. But, because evaluating f (a) is easy, in that case one will recognize the mistake and start over.[0396] An unstructured search problem, in the case when the initial state is unknown and arbitrary entangled is the most general situation one might expect to have when working with subroutines involving quantum search or counting of solutions in a larger quantum computation. This situation is typical in the case of KB design of robust fuzzy controllers in intelligent control suspension system for different types of roads and connected with partial sorted data after GA optimization and an FNN learning processes. In particular, it is useful to derive an iteration formula for the action of the original Grover's operator and find, similar to the case of an initial state with unknown amplitudes, that the final state is a periodic function of the number of “good” items and can be expressed in terms of first and second order moments of the initial amplitude distribution of states alone. [0397] Considered the problem to find a “good” file, represented as the state |g , out of N files |a; a=0, . . . , N−1. The algorithm starts with the preparation of a flat superposition of all states |a, i.e.[0398] and assumes that there is an oracle which evaluates the function H(a), such that H(g) 1 for the “good” state |g , and H(g)=0 for the “bad” states |b (i.e., the remaining states in the set of all the a's). The unitary transformation for the “search” of |g is then defined by G_{H}=−WS_{O}WS_{H}, where the Walsh-Hadamard transform W is defined as
[0399] (with
[0400] a [0401] In fact S _{0} with an extra ancillary qubit |e≅[|0−|1]/{square root}{square root over (2)}, such that U_{H}|a|e≅|a|e+H(a)mod2).
[0402] Thus obtaining
[0403] Iterating G [0404] New algorithms with exponential speed-up in Appendix 5 are described. [0405] The algorithms discussed above made the essential assumption that the starting state is to be prepared in the flat superposition form given by Equation (8.20). A first attempt to generalize such results occurs when the amplitudes of the initial superposition of states are arbitrary and unknown complex numbers. In particular, by exactly solving certain linear differential equations describing the evolution of the initial amplitudes, one can still express the optimal measurement time and the maximal probability of success in a closed and exact form which depends only on the averages and the variances of the initial amplitude distribution of states. [0406] One of the main resources and ingredients of quantum computation lies, however, not only in the possibility of dealing with arbitrary complex superpositions of qubits, but in the massive exploitation of quantum entanglement. One cannot necessarily deal with the ansatz when the “good” state has a complicated, unknown structure by entanglements by directly and naively using Grover's algorithm. An important case may arise, for instance, when the computational qubits get nontrivially entangled with environment, and encoding/decoding techniques become necessary in order to prevent errors from occurring and spreading in the quantum computer calculations. Different approaches for solving these cases in Appendix 5 are discussed. [0407] Consider, for example, if in the database search problem, one is given the initial superposition
[0408] where now the index a simply labels the files, while f(a) corresponds to the actual file content. In fact, one might know the desired states |g , but ignore the function f (and, therefore, the file content f(g)), and thus want to extract the states|g|f(g)) from the original superposition and eventually read (i.e. measure) or use f(g) only later in another quantum routine.[0409] But, in Grover's algorithm, the application of any unitary transformation acting on the label states |a would automatically affect also and nontrivially (e.g., producing complicated entangled states) on |f(a)), with f(a) unknown a priori. Grover's algorithm is generalized, for an arbitrary entangled initial state, by giving an exact formula for the n-th iteration of Grover's operator and comparing the results with those for the case of an initial superposition of states with arbitrary complex amplitudes.[0410] The “good” (orthonormal) states to be found are defined, in number t, as |g , the remaining, or “bad” states are defined as |b, where, by definition,[0411] Then study what is the effect of acting with Grover's unitary transformation G [0412] where ≡|G _{2} . By induction, the n-th iteration of G_{H}on |ψ gives
[0413] where the states |X _{2} ^{(n)} satisfy the following recurrence relations
[0414] with |C _{2} +(−1)^{n}|B_{2} and sin^{2 }θ=t/N, and are subject to the initial condition |X_{2} ^{(1)} =|Y_{2} ^{(1)}=.
[0415] Adopting a more compact matrix notation, i.e. writing {right arrow over (Z)} _{n},|Y_{2} ^{(n)} →Y_{n }and |C_{2} ^{(n)} →C_{n }and defining the matrices M≡cos 2θI+M_{1 }with M_{1}≡cos 2θσ_{x}+iσ_{y }the recurrence equations (8.25) subject to the initial condition X_{1}=Y_{1}=C_{1 }can be transformed into the simple matrix equation
[0416] Equation (8.26) can be solved using standard techniques to give
[0417] (with [k] being the integer part of k), and where the n-th powers of the matrices M and M [0418] Inserting equation (8.28) into equation (8.27), one obtains
[0419] and, finally, from equation (8.24), the formula for the n-th iteration of G [0420] where ^{n }tan θ(tan nθ)^{(−1} ^{ n }|B_{2} )
[0421] Similar to the case of the original Grover's algorithm acting on an initial flat superposition of states, G n+1)θ]|ω+cos[(2n+1)θ]|r (8.31)
[0422] where |ω ≡|G_{1} /{square root}{square root over (t)} and |r≡B_{1} /{square root}{square root over (N−t)}.
[0423] A general normalization is given by
[0424] and substituting for [0425] |f′(g) ≡f(g)/{square root}{square root over (N′)}, |G′≡|G/{square root}{square root over (N′)}, |G_{2}′|G_{2} /{square root}{square root over (N′)} (and similarly, substituting everywhere g→b, for |f′(b), |B′ and |B_{2}′), one can write the initial normalized and entangled state as |ψ≡|G′+|B′, and rewrite equation (8.30) as
[0426] where the quantities
[0427] have been introduced and, by definition, |{overscore (G)} _{G}|G_{2}′, |{overscore (B)}_{2}′^{(0)} ≡γ_{B}|B_{2}′ with γ_{G}≡1/t and γ_{B}≡1/(N−t). Further defining the averages and variances
[0428] and similarly for |{overscore (B)} _{2}′ and σ_{B} ^{2(n) }(after the substitution g→b), one can easily show that
[0429] which, inserted into Eqs.(8.33), give
[0430] and, finally, from Eqs.(8.36) one finds the constants of the motion.
[0431] Defining the quantities
[0432] and introducing the angle ω≡2θ, makes it possible to rewrite the norms of the states (8.37) as
[0433] where φ≡φ [0434] and similarly for σ [0435] can be finally written, using equations (8.36),(8.39) and (8.40), as
[0436] The probability P(n) can be found to be maximized, P _{max}=1) either provided that t=N (trivial case) or that the following conditions on the moment of the amplitudes of the initial distribution of states are satisfied
φ [0437] i.e., for [Re {overscore (G)}_{2}′^{(0)}|{overscore (B)}_{2}′_{(0)} ]^{2}={overscore (G)}_{2}′^{(0)}|{overscore (G)}_{2}′^{(0)} {overscore (G)}_{2}′^{(0)}|{overscore (G)}_{2}′^{(0)} (which can be true, e.g., if |{overscore (G)}_{2}′^{(0)} =c|{overscore (G)}_{2}′^{(0)} with cεR) and for γ_{B} B′|B′={overscore (G)}_{2}′^{(0)}|{overscore (G)}_{2}′^{(0)} . In particular, for n=n_{j},
[0438] Of course, the unitary nature of the operator prevents one from naively getting only the exact contribution from the initial “unperturbed” entangled states |f′(g) in G_{H} ^{nj}|{overscore (ψ)}. However, as some elementary algebra can show, it is still possible, for instance in the case of a large enough number of “good” items g, i.e. for t/N ≦O(1), to make the amplitude contribution coming from the other entangled states |{overscore (G)}_{2}′^{(0)} and {overscore (B)}_{2}′^{(0)} relatively small compared to that of |f′g), if j is even and provided that {overscore (G)}_{2}′^{(0)}|{overscore (G)}_{2}′^{(0)} /{overscore (B)}_{2}′^{(0)} |{overscore (B)}_{2}′^{(0)} {overscore (>)}O(1).
[0439] Finally, it is also straightforward to show that the particular case of an initial state with arbitrarily complex amplitudes can be recovered provided one makes the substitutions |f(g) →{square root}{square root over (N)}k_{i},|f(b)→{square root}{square root over (N)}I_{j}, t→r and n→t, with k_{i }and l_{i }complex numbers. The maximum probability of success P_{max }can be achieved again after n_{j }steps, and corresponds to certainty if one has
[0440] i.e. when l [0441] The algorithm COUNT, described below, is used for the case of an initial flat superposition of states. The COUNT algorithm essentially exploits Grover's unitary operation G [0442] The COUNT algorithm involves the following sequence of operations:
[0443] Since the amplitude of the set of the good states |g after m iterations of G_{H }on |a is a periodic function of m, the estimate of such a period by use of the Fourier analysis and the measurement of the ancilla qubit |m will give information on the size t of this set, on which the period itself depends. The parameter P determines both the precision of the estimate t and the computational complexity of the COUNT algorithm (which requires P) iterations of G_{H}.
[0444] Using the more general normalization
[0445] (with 0<N [0446] Then act on |{overscore (ψ)} in equation (8.45) with an |m—“controlled” Grover operation G_{H} ^{m }and on |m with a Fourier transform F, thus getting
[0447] As in the standard COUNT algorithm, requiring that the time needed to compute the repeated Grover operations G [0448] Summing over n in equation (8.46), after some elementary algebra gives (taking, without loss of generality, P even)
[0449] where the following quantities have been introduced:
[0450] and where the states |A , |B, |C are mutually orthogonal and given by[0451] At this point one can rewrite formula (8.47) in the general case when f is not an integer, distinguishing three possible cases. In particular, when 0<f<1 |{overscore ( _{1} +|1 |b _{1} +|P−1 |c _{1} +|R _{1} (8.50)
[0452] where |R 104 )}_{3} not containing the ancillary qubits |0, |1, |P−1. One can show that the total probability amplitude in the first three terms (i.e., the probability that, in a measurement of the first ancillary qubit, one obtains any of the states |0, |1, |P−1) is given by
[0453] with
[0454] and it can be shown that
[0455] When 1<f<P/2, instead, |{overscore (ψ)} _{2} +|P/2|b _{2} +P/2+1|c _{2} +R _{2} (8.52)
[0456] where the meaning of |R [0457] with
[0458] and it can be shown that
[0459] Finally, in the most general case in which 1<f<P/2−1 |{overscore (ψ)} _{3} +|P−f ^{−} |b _{3} +|f ^{+} |c _{3} +|P−f ^{+} |d _{3} +R _{3} (8.54)
[0460] where |R ^{−}≡[f]+δf and f^{+}≡f^{−}+1 with 0<δf<1.
[0461] The total probability amplitude in the first four terms in this case is given by
[0462] and again
[0463] The final step of the COUNT algorithm involves measuring the first ancillary qubit in the state |{overscore (ψ)} _{±} , |P−f_{±} , respectively, for the three cases, and, therefore, still be able to evaluate the number t of “good” states from sin θ={square root}{square root over (t/N)} and equation (8.48) even in the case of an initial entangled state with the same probability as in the case of an initial flat superposition of states, it is desirable to impose the condition
[0464] The probability can be made exponentially close to one by repeating the whole algorithm many times and using the majority rule. The probabilities W _{R} and then acting with a |m_{1} . . . |m_{R} —“controlled” G_{H} ^{m }operation on the state |{overscore (ψ)}.)
[0465] Taking for simplicity N=N′ in Equation (8.55) for general case 1<f<P/2-1, Equation (8.56) would lead to the condition on the initial averages
[0466] which, for example, upon the choice |{overscore (G)} _{2}′^{(0)} , would require that c^{2}>(2−1/γ_{B})γ_{G}. Furthermore, since in general f is not an integer, the measured {tilde over (f)} will not match exactly the true value of f and, defining {tilde over (t)}≡N sin^{2 }{tilde over (θ)}, with {tilde over (θ)}={tilde over (θ)}({tilde over (f)}), gives, for the error over t, the same estimate, i.e.
[0467] so that the accuracy will always remain similar in the cases of an initial unentangled or entangled state. [0468] In the most general case when Grover's algorithm is to be used as a subroutine in a bigger quantum network when the generic form of the initial state is unknown and arbitrary entangled superposition of qubits. In particular, one can preserve a good success probability and a high accuracy in determining the number of “good” items even if the initial state is entangled, again provided that some conditions are satisfied by the averages and variances of the amplitude distribution of the initial state. [0469] Consider the situation where the number of objects satisfying the search criterion is greater than 1. Let a database {w [0470] Here the l elements {w _{j} |j=1, . . . , N}. Let Λ=span{|w_{j} |1≦j≦1} be the subspace of spanned by the vectors of the good objects. (To avoid introducing another layer of subscripts, it is assumed that these good objects are the first l items.)
[0471] Now, define a linear operation in terms of the oracle function as follows: j=1, 2, . . . , N. (9.1)
[0472] Then since I [0473] where I is the identity operator on . I_{Λ} is the operator of rotation (by π) of the phase of the subspace Λ.
[0474] The explicitness of (9.2) is misleading because explicit knowledge of {|w |≦j≦l} in (9.2) is not available. Nevertheless, (9.2) is a well-defined (and unitary) operator on because of (9.1).
1 [0475] Now again define |s as[0476] where now
[0477] As before, T _{s }is unitary and hence quantum-mechanically admissible. I_{s }is explicitly known, constructible with the so-called Walsh-Hadamard transformation.
[0478] Let {tilde over (Λ)}=span{Λ∪|r }. Then {|w_{i} , |r|i=1, 2, . . . , l} forms an orthonormal basis of {tilde over (Λ)}. The orthogonal direct sum H={tilde over (Λ)}⊕{tilde over (Λ)}^{⊥} is an orthogonal invariant decomposition for both operators I_{{tilde over (Λ)}} and I_{s}. The restriction of I_{s }of {tilde over (Λ)}^{⊥} is P_{{tilde over (Λ)}⊥}, the orthogonal projection operator onto {tilde over (Λ)}^{⊥}. From (9.3),
[0479] Furthermore, the conclusion follows: 1) the restriction of I [0480] Consequently, I [0481] The generalized Grover search engine for multi-object search is now constructed as: [0482] Substituting (9.2) and (9.4) into (9.7) and simplifying,
[0483] The orthogonal direct sum H={tilde over (Λ)}⊕ _{2} , . . . , |w_{1} , |r} of {tilde over (Λ)}, the operator U admits the real unitary matrix representation
[0484] 2) The restriction of U to {tilde over (Λ)} [0485] The results above effect a reduction of the problem to an invariant subspace {tilde over (Λ)}. However, {tilde over (Λ)} is a (l+1)—dimensional subspace where l may be fairly large. Another reduction of dimensionality is needed to further simplify the operator U. [0486] Define by[0487] Let
[0488] Then {|{tilde over (w)} , |r} forms an orthonormal basis of . Then is an invariant two-dimensional subspace of U such that:[0489] 1) r, sε ; 2) U()=.[0490] One has the second reduction, to dimensionality 2. [0491] Using matrix representation (9.8) and (9.9), and the definition of |{tilde over (w)} as[0492] one obtains the following: With respect to the orthonormal basis {|{tilde over (w)} , |r} in the invariant subspace , U admits the real unitary matrix representation[0493] Since |s ε, one can calculate U^{m}|s using (9.10):
[0494] Thus, the probability of reaching the state |{tilde over (w)} after m iterations is [0495] If l<<N, then α is close to
[0496] and, therefore, equation (9.12) is an increasing function of m initially. This again manifests the notion of amplitude amplification. This probability P [0497] When
[0498] is small
[0499] The generalized Grover algorithm for multi-objective search with operator U given by (9.7) has the success probability P [0500] small, after
[0501] iterations, the probability of reaching |{tilde over (w)} εΛ is close to 1.[0502] The result (9.13) is consistent with Grover's original algorithm for single object search with l=1, which has
[0503] Assume that
[0504] is small. Then, any search algorithm for l objects, in the form of U _{j}, j=1, 2, . . . , p is an unitary operator and |w_{l}) is an arbitrary superposition state, takes in average
[0505] iterations in order to reach the subspace Λ with a positive probability
[0506] independent of N and l. [0507] Unfortunately, if the number l of good items is not known in advance, the above does not show when to stop the iteration. Consider stopping the Grover process after j iterations, and, if a good object is not obtained, starting it over again from the beginning. The probability of success after j iteration is cos 0 [0508] Now approximate the solution j of (9.14) iteratively as follows. The first order approximation j [0509] Higher order approximations j [0510] based on equation (9.14). This process will yield a convergent solution j to (9.14). Information analysis of this problem in Appendix 4 is developed. [0511] In FIG. 1, the GA 131 produces an optimal solution from single space of solution. The GA 131 compresses the value information from a single solution space with the guarantee of the safety of informative parameters in general signal K of the PID controller 150. [0512] In FIG. 20 the GA 131 produces a number of solutions as structured (sorted) data for the QGSA 2001. The quantum search algorithm on structured (sorted) data searches for a successful solution with higher probability and greater accuracy than a search on unstructured data. The input to the QGSA 2001 is a set of vectors (string) and the output of the QGSA 2001 is a single vector K. A linear superposition of cells of look-up tables of fuzzy controllers in the QGSA 2001 is produced with the Hadamard Transform H. Components of the vector K are coded as qubits, either |0 or |1. The Hadamard transform H is formed independent for every qubit a linear superposition of qubits.[0513] For example, consider a qubit
[0514] With a unitary matrix as a Hadamard transform
[0515] Thus
[0516] The QGSA 2001 evolves classical states as cells of look-up tables from the GA 131 or for the FNN 142 into a superposition and therefore cannot be regarded as classical. The collection of qubits is a quantum register. This leads to the tensor product (product in Hilbert space). The tensor product is identified with the Kronecker product of matrices. The next step involves coding of information. As in the classical case, it can be used to encode more complicated information. For example, the binary form of 9 (decimal) is 1001 and after loading a quantum register with this value is done by preparing four qubits in state |9 ≡|1001≡|1 |0 |0 |1. Consider first the case with two qubits. With the basis |00≡|0 |0, |01≡|0 |1, |10≡|1 |0, |11≡|1 |1. If one initialises a quantum memory register so that in the register starts out in the |0z,900 ) state then applies a Hadamard gate to each qubit independently, the net result places the entire n-qubit register in a superposition of all possible bit strings that an n-bit classical register cannot. Thus, using the Hadamard gate, one can effectively enter 2^{n }bit strings into a quantum memory register using only n basic operations.
[0517] Applying qubits individually, one can obtain the superposition of the 2 [0518] Thus one can effectively load exponentially (i.e., 2 [0519] In the general case of the PID controller 150, K(t)={k _{i l (t). }
[0520] The tensor product operations: |10 ≡|1 |0 means that the logic joint of signal states, as example, between k^{i} _{1}(t) and k^{i} _{2}(t) is given for a PID controller. According to the SSCQ 130 the vector tensor product describes the joint probability amplitude of two systems of being in a joint state. The random optimal output of the GA is the single vector K with stochastically independent components k_{i}(t).
[0521] Using a Grover-type quantum search algorithm one can realise more simple robust control with the co-ordination of signals k [0522] With this method one can check the robustness of the look-up table as a the knowledge base for the fuzzy PID controller. Grover's quantum algorithm is a tool for searching for a solution as one universal robust look-up table from many look-up tables of fuizzy controllers for an intelligent smart suspension control system. [0523] Consider, for example, the case n=2 and x=01 [0524] f(00)=0, f(01)=1, f(10)=0, f(11)=0. [0525] in order to study the robustness of one cell in one look-up table for a fuzzy controller. [0526] The entanglement operator is:
[0527] and |input =|00 |1. An entanglement operator defines a permutation of basis vectors of the superposition it is applied, mapping one basis vector into another basis vector, but not into a superposition. By applying the superposition, the entanglement and the interference operator in sequence we obtain the final vector (see, Appendix 1)[0528] Reading the value of the first two qubits after simulation of suspension system stochastic behavior the searched state x is found. [0529] Temporal labeling is used to obtain the signal from the pure initial state
[0530] by repeating the simulation experiment three times, cyclically permuting the |01 , |10 and |11 state populations before the computation and then summing the results. The calculation starts with a Walsh-Hadamard transform W (H), which rotates each quantum bit (qubit) from |0 to (|0+|1)/{square root}{square root over (2)}, to prepare the uniform superposition state[0531] From physical standpoint W=H _{B}, where H=X^{2}{overscore (Y)} (pulses applied from right to left) is a single-spin Hadamard transformation. These rotations are denoted as X≡exp(iπI_{x}/2) for a 90° rotation about {circumflex over (x)} axis, and Y≡exp(iπI_{y}/2) for a 90° rotation about ŷ axis, with a subscript specifying the affected spin. The operator corresponding to the application of f(x) for x_{0}=3 is a
[0532] This conditional sign flip, testing for a Boolean string that satisfies the AND function, is implemented by using the coupled-spin evolution. During a time t the system undergoes the unitary transformation exp(2πiJI [0533] An arbitrary logical function can be tested by a network of controlled-NOT and rotation gates, leaving the result in a scratch pad qubit. This qubit can then be used as the source for a controlled phase-shift gate to implement the conditional sign flip. [0534] The operator D in Grover's quantum search algorithm that inverts the states about their mean can be implemented by a Walsh-Hadamard transform W, a conditional phase shift P, and another Was following
[0535] Let U DC be the complete iteration. The state after one cycle is
[0536] Measurements of the system's state will give with certainty the correct answer, |11 .[0537] For further iterations, |ψ ^{n}|ψ_{0} ,
[0538] A maximum in the amplitude of the x [0539] In one embodiment, like any computer program that is compiled to a micro-code, the Radio Frequency (RF)—pulse sequence for U can be optimized to eliminate unnecessary operations. In a quantum computer this is desirable to make the best of the available coherence. Ignoring irrelevant overall phase factors, and noting that H=X [0540] ρ _{n}|−tr(|ψ_{n} ψ_{n}|)/4,
[0541] to be measured. [0542] The effect of the elementary rotation G is shown in FIG. 24 for the case of three qubits, i.e. m=3. The first Hadamard transformation H _{s }this gate sequence G amplifies the probability amplitude of the searched state |x_{0} =|111. In this particular case an additional Hadamard transformation finally prepares the quantum computation in the searched state |x_{0} =|111 with a probability of 0.88. This method for global optimization and design of KB in fuzzy (P)(I)(D)-controllers is used.
[0543] The main application problem of quantum search algorithm in optimization of fuzzy controller KB is the increasing of memory size in simulation on classical computer. An algorithm for this case is provided in Appendix 3, and an example of this the use of this algorithm is described below. [0544] An example for a set of binary patterns of length 2 will help clarify the preceding discussion. Assume that the pattern set for fuzzy P-controller is p={01, 10, 11}. Recall (from Appendix 3) that the x register is the one that corresponds to the various patterns, that the g register is used as a temporary workspace to mark certain states and that the c-register is a control register that is used to determine which states are affected by a particular operator. Now the initial state |00, 0, 00) is generated and the algorithm evolves the quantum state through the series of unitary operations. [0545] First, for any state whose c _{1 }qubit's state is flipped if the c_{2 }qubit is |0. This flipping of the c_{1 }qubit's state marks this state for being operated upon by an Ŝ^{p }operator in the next step. So far, there is only one state, the initial one, in the superposition. This flipping is accomplished with the FLIP operator: |00, 0, 00→^{FLIP}|01, 0, 10
[0546] Next, one state in the superposition with the c—register in the state |10 (and there will always be only one such state at this step) is operated upon by the appropriate SP operator (with p equal to the number of patterns including the current one yet to be processed, in this case 3). This essentially “carves off” a small piece and creates a new state in the superposition. This operation corresponds to[0547] Next, the two states affected by the SP operator are processed by the SAVE operator of the algorithm. This makes the state with the smaller coefficient a permanent representation of the pattern being processed and resets the other to generate a new state for the next pattern. At this point one pass through the loop of the algorithm has been performed.
[0548] Now, the entire process is repeated for the second pattern. Again, the x register of the appropriate state (that state whose c [0549] Next, another Sp operator is applied to generate a representative state for the new pattern:
[0550] Again, the two states just affected by the Ŝ [0551] Finally, the third pattern is considered and the process is repeated a third time. The x register of the generator state is again selectively flipped. In this time, only those qubits corresponding to bits that differ in the second and third patterns are flipped, in this case just qubit x [0552] Again a new state is generated to represent this third pattern.
[0553] Finally, proceed once again with the SAVE operation.
[0554] At this point, notice that the states of the g and c registers for all the states in the superposition are the same. This means that these registers are in no way entangled with the x register, and therefore since they no longer needed they may be ignored without affecting the outcome of father operations on the x register. Thus, the simplified representation of the quantum state of the system is
[0555] and it may be seen that the set of patterns p is now represented as a quantum superposition in the x register. [0556] In the quantum network representation of the algorithm the FLIP operator is composed of the {circumflex over (F)} [0557] In looking for the state |0110 , assume that the first two steps of the algorithm (which initialize the system to the uniform distribution) have not been performed, but that instead initial state is described by[0558] that is superposition of only 6 of the possible 16 basis states. The first time through the loop, inverts the phase of the state |τ =|0110 resulting in[0559] and then rotates all the basis states about the average, which is
[0560] so
[0561] The second time through the loop, again rotates the phase of the desired state giving
[0562] and then again rotates all the basis states about the average which now is
[0563] so that
[0564] Now squaring the coefficients gives the probability of collapsing into the corresponding state. In this case, the chance of collapsing into the |τ =|0110 basis state is 0.66^{2}≈44%.
[0565] The chance of collapsing into one of the 15 basis states that is not desired state is approximately 56%. This chance of success is much worse than that seen in above described example, and the reason for this is that there are now two types of undesirable states: those that existed in the superposition to start with but that are not the state we are looking for and those that were not in the original superposition but were introduced into the superposition by the Ĝ operator. The problem comes from the fact that these two types of undesirable states acquire opposite phases and thus to some extent cancel each other out. Therefore, during the rotation about average performed by the Ĝ operator the average is smaller than it should be if it were to just represent the states in the original superposition. As a result, the desired state is rotated about a sub-optimal average and never gets as large a probability associated with it as it should. An analytic expression for the maximum possible probability using Grover's algorithm on an arbitrary starting distribution is
[0566] where N is the total number of basis states, r is the number of desired states (looking for more than one state is another extension to the original algorithm), l [0567] Now consider the case of the initial distribution. The variance is proportional to 10·0.13 [0568] The number of strings in a population matching (or belonging to) a schema is expected to vary from one generation to the next according to the following theorem:
[0569] where m (H,t) is the number of strings matching the schema H at generation t, f(H,t) is the mean fitness of the strings matching H, {overscore (f)}(t) is the mean fitness of the strings in the population, p [0570] As stated above, the GA searches for a global optimum in a single solution space. It is desirable, however, to search for a global optimum in multiple solution spaces to find a “universal” optimum. A Quantum Genetic Search Algorithm (QGSA) provides the ability to search multiple spaces simultaneously (as described below). The QGSA searches several solution spaces, simultaneously, in order to find a universal optimum, that is, a solution that is optimal considering all solution spaces. [0571] The structure of quantum search algorithm can be described as
[0572] In quantum algorithm structures and genetic algorithms structure have the following interrelations:
[0573]FIG. 25 illustrates the similarities between a GA and a QSA. As shown in FIG. 25, in the GA search, a solution space [0574] By contrast, in the QSA shown in FIG. 25, a group of N solution spaces [0575] Thus, the classical process of selection is loosely analogous to the quantum process of creating a superposition. The classical process of crossover is loosely analogous to the quantum process of entanglement. The classical process of mutation is loosely analogous to the quantum process of interference. [0576] In the GA a starting population is randomly generated. Mutation and crossover operators are then applied in order to change the genome of some individuals and create some new genomes. Some individuals are then cut off according to a fitness function and selection of good individuals is used to generate a new population. The procedure is then repeated on this new population until an optimum is found. [0577] By analogy, in the QSA an initial basis vector is transformed into a linear superposition of basis vector by the superposition operator. Quantum operators such as entanglement and interference then act on this superposition of states generating a new superposition where some states (the non-interesting states) have reduced their probability amplitude in modulus and some other states (the most interesting) have increased probability amplitude. The process is repeated several times in order to get to a final probability distribution where an optimum can be easily observed. [0578] The quantum entanglement operator acts in analogy to the genetic mutation operator: in fact it maps every basis vector in the entering superposition into another basis vector by flipping some bits in the ket label. The quantum interference operator acts like the genetic crossover operator by building a new superposition of basis states from the interaction of the probability amplitudes of the states in the entering superposition. But the interference operator includes also the selection operator. In fact, interference increases the probability amplitude modulus of some basis states and decreases the probability amplitude modulus of some other ones according to a general principle, that is maximizing the quantity
[0579] with T={1, . . . , n} This quantity is called the intelligence of the output state and it measures how the information encoded into quantum correlation by entanglement is accessible by measurement. The role of the interference operator is, in fact, to preserve the Von Neumann entropy of the entering entangled state and to reduce (minimize) the Shannon entropy, which has been increased to its maximum by the superposition operator. Note that there is a strong difference between GA and QSA: in GA the fitness functions changes with different instances of the same problem, whereas mutation and crossover are always random. In the QSA, the fitness function is always the same (the intelligence of the output state), whereas the entanglement operator strongly depends on the input function f. [0580] The QGSA merges the two schemes of GA and QSA. FIG. 26 is a flowchart showing the structure of the QGSA. In FIG. 26, an initial superposition with t random non-null probability amplitude values is generated
[0581] Every ket corresponds to an individual of the population and in the general case is labelled by a real number. So, every individual corresponds to a real number x [0582] The entanglement operator includes an infective map transforming each basis vector into another basis vector. This is done by defining a mutation ray ε>0 and extracting t different values ε [0583] When U [0584] The mutation operator ε can be described as following relation
[0585] Assume, for example, there are eight states in the system, encoded in binary as 000, 001, 010, 011, 100, 110, 111. One of the possible states that may be found during a computation is
[0586] A unitary transform is usually constructed so that it is performed at the bit level. For example, the unitary transformation
[0587] will switch the state |0 to |1 and |1 to |0 (NOT operator).[0588] Mutation of a chromosome in the GA alters one or more genes. It can also be described by changing the bit at a certain position or positions. Switching the bit can be simply carried out by the unitary NOT-transform. The unitary transformation that acts, as example on the last two bits will transform the state |1001 to state |1011 and the state |0111 to the state |0101 and so on can be described as the following matrix[0589] which is a mutation operator for the set of vectors |0000 , |0001, . . . , |1111.[0590] A phase shift operator Z can be described as the following
[0591] and an operator
[0592] is a combination of negation NOT and a phase shift operator Z. [0593] As an example, consider the following matrix
[0594] which operates a crossover on the last two bits transforming 1011 and 0110 in 1010 and 0111, where the cutting point is at the middle (one-point crossover). [0595] The two-bit conditional phase shift gate has the following matrix form
[0596] and the controlled NOT (CNOT) gate that can created entangled states is described by the following matrix:
[0597] The interference operator Int [0598] The average entropy value for this state is now evaluated. Let E(x) be the entropy value for individual x. Then
[0599] The average entropy value is calculated by averaging every entropy value in the superposition with respect to the squared modulus of the probability amplitudes. [0600] According to this sequence of operations, k different superpositions are generated from the initial one using different entanglement and interference operators. Every time the average entropy value is evaluated. Selection involves keeping only the superposition with minimum average entropy value. When this superposition is obtained, it becomes the new input superposition and the process starts again. The interference operator that has generated the minimum entropy superposition is kept and Int [0601] with x [0602] Int [0603] with −ε≦ε [0604] with Int [0605] 6. If {overscore (E)}*<E [0606] 7. Else set |input to |output *, Int^{1 }to Int^{j* }(block 2608) and go back to step 2 (block 2602).
[0607] Step 6 includes methods of accuracy estimation and reliability measurements of the successful result. [0608] The simulation of the quantum search algorithm is represented through information flow analysis, information risk increments and entropy level estimations: [0609] 1) Applying a quantum gate G on the input vector stores information into the system state, minimizing the gap between the classical Shannon entropy and the quantum Von Neumann entropy; [0610] 2) Repeating the step of applying the calculation (estimation) of information risk increments (see below); [0611] 3) Measuring the basis vector for estimation of the level of the average entropy value; [0612] 4) Decoding the basis vector of a successful result for computation time stopping when the minimum average entropy value falls under a given critical level limit. [0613] The information risk increments are calculated (estimated) according to the following formula: −{square root}{square root over ( [0614] where: [0615] W is the loss function; [0616] r(W [0617] x=(x [0618] θ is an unknown parameter;
[0619] is the relative entropy (the Kullback-Leibler measure of information divergence). [0620] As stated above, the GA searches for a global optimum in a single solution space. As shown in FIG. 25, in the GA search, a solution space [0621] The “single solution space” can include coefficient gains of the PID controller of a plant under stochastic disturbance with fixed statistical properties as the correlation function and probability density function. After stochastic simulation of dynamic behaviour of the plant under stochastic excitation with the GA one can obtain the optimal coefficient gains of intelligent PID controller only for stochastic excitation with fixed statistical characteristics. In this case the “single space of possible solutions” is the space 2501. Using a stochastic excitation on the plant, with another statistical characteristics, then the intelligent PID controller can not realize a control law with the fixed KB. In this case, a new space of possible solutions, shown as the space 2550, is defined. [0622] If a universal look-up table for the intelligent PID controller is to be found from many single solution spaces, then the application of the GA does not give a final corrected result (the GA operators not include superposition and quantum correlation as entanglement). The GA gives the global optimum on the single solution space. In this case important information about statistical correlation between coefficient gains in the universal look-up table is lost. [0623] By contrast, in the QSA shown in FIG. 25, a group of N solution spaces [0624] The structure of intelligent suspension control system is shown in FIG. 21. FIG. 33 shows a look-up table fragment simulation for the fuzzy P-controller by the GA of FIG. 21. This example shows the application of the QGSA for the optimization of a look-up table for the P-controller of a suspension system using two look-up tables. The two look-up tables from GA simulations for Gaussian and non-Gaussian (with Rayleigh probability density function) roads corresponding to the road profiles in FIGS. 4 and 6. [0625] Stepper motors of dampers in the suspension system make the positions from the discrete interval [1, 2, . . . , 9]. In this example, there is a relation between the error control (ε) and the change of error control ({dot over (ε)}) as [PM→NB] for the different position states of two dampers. The two look-up tables cannot be simply averaged together. Only with a quantum approach using superposition operator the Cell [0626] Assume, for example, the selection operator of the GA codes randomly the position of a damper in the Cell [0627] With the modification of the quantum search algorithm described above
[0628] The first two steps are identical to those above:
[0629] Now, all the states present in the original superposition are phase rotated and then all states are again rotated about average:
[0630] Squaring the coefficients gives the probability of collapsing into the desired |τ =|0110 basis state as 99%—a significant improvement that is critical for the Quantum Associative Memory described in the next section.[0631] Quantum Associative Memory Structure [0632] A quantum associative memory (QuAM) can now be constructed from these algorithms. Define {circumflex over (P)} as an operator that implements the algorithm for memorizing patterns described above. Then the operation of the QuAM can be described as follows. Memorizing a set of patterns is simply |ψ ={circumflex over (P)}|{circumflex over (0)},[0633] with |ψ being a quantum superposition of basis states, one for each pattern. Now, assume n−1 bits of a pattern are known and the goal is to recall the entire pattern. The modified Grover's algorithm can be used to recall the pattern as[0634] followed by |ψ ={circumflex over (GI)}|ψ,[0635] repeated T times (how to calculate T is covered in Appendix 3 and below), where τ=b [0636] As an example of the QuAM, assume a set of patterns p={0000,0011,0110,1001,1100,1111} is known. Then using the notation of the above described example, a quantum state that stores the pattern set is created as
[0637] Now assume that the pattern whose first three bits are 011 is to be recalled. Then τ=011?, and applying this equation gives
[0638] At his point, there is a 96.3% probability of observing the system and finding the state |011 ?). Of course there are two states that match and state |0111) has a 22% chance. This may be resolved by a standard voting scheme. Observation of the system shows that the completion of the pattern 011 is 0110. [0639] Dynamic analysis is described here and in Appendix 4. The results of information analysis, together with dynamic evolution of quantum gate for Grover's algorithm, begins by considering the operator that encoding the input function as:
[0640]FIG. 34 shows a general iteration algorithm for information analysis of Grover's QA. In FIGS. 35 and 36 two iterations of this algorithm are reported. From these figures it is observed that: [0641] 1. The entanglement operator in each iteration increases correlation among the different qubits; [0642] 2. The interference operator reduces the classical entropy but, as side effect, it destroys part of the quantum correlation measure by the Von Neumann entropy. [0643] Grover algorithm builds intelligent states in several iterations. Every iteration first encodes the searched function by entanglement, but then partly destroys the encoded information by the interference operator. Several iterations are needed in order to conceal both the need to have encoded information and the need to access it. The Principle of Minimum Classical (Quantum) Entropy in the output of QA leads to a successful result on intelligent output states. The searching QA's (such as Grover's algorithm) check for minimum of Classical Entropy and co-ordination of the gap with Quantum Entropy Amount. The ability of co-ordination of these two values characterises the intelligence of searching QA's. [0644] When the output vector from the quantum gate has been measured, it must interpret it in order to find x>|1> with probability near to 1. The output vector is encoded back into binary values using the first n basis vector in the resulting tensor product, obtaining string x as the final answer.
[0645] For example, assume that n=2 [0646] A probability of success search can be developed by letting N be the total number of basis states, r [0647] k [0648] k [0649] l [0650] l [0651] Here k [0652] A little more algebra gives the averages as
[0653] Now consider this new state described by these equations as the arbitrary initial distribution to which the results can be applied. These can be used to calculate the upper bound on the accuracy of the QuAM as well as the appropriate number of times to apply this equation in order to be as close to that upper bound as possible. The upper bound on accuracy is given by [0654] whereas the actual probability at a given time t is [0655] The first integer time step T for which the actual probability will be closest to this upper bound is given by rounding the function
[0656] to the nearest integer. [0657] The algorithm described above can handle only binary patterns. Nominal data with more than two values can be handled by converting the multiple values into a binary representation. 11. Quantum Optimization, Quantum Learning and Robustness of the Fuzzy Intelligent Controller [0658] One embodiment includes extraction of knowledge from the simulation results and forming a robust Knowledge Base (KB) for the fuzzy controller in the intelligent suspension control system (ISCS). The basises for this approach are Grover's QSA (optimization of unified look-up table structure) and quantum learning (KB production rules with relatively minimal sensitivity to different random excitations of the control object). [0659] According to the structure in FIG. 20, consider the summarization role of Grover's QSA in the process of forming the teaching signal for the KB fuzzy controller. Appendices 2, 3 and 4 provide further descriptions of Grover's QSA operations and model structures. [0660] 11.1. Standard Grover's QSA structure and Results of the Measurement Process. The individual outcomes of a measurement process can be understood within standard quantum mechanics in terms of executing Grover's QSA. A measurement interaction first entangles system S with the measuring process X. In general, one obtains the state
[0661] where the states |X _{i} with a definite apparatus state |X_{i} . Since, this is an entangled state, it must be reduced to a particular state |S_{i} |X_{i} before the result can be read off. This is achieved by a non-unitary process by projecting the state |ψ to this state with the help of the projection operator Π_{i}=|X_{i} X_{i}|. One can obtain the reduced density matrix
[0662] which is diagonal and represents a heterogeneous mixture with probabilities lc. 2. [0663] The algorithm amplifies the amplitude of an identified target (the amplitude corresponding to a particular eigenstate in this case) at the cost of all other amplitudes to a point where the latter becomes so small that they cannot be recorded by detectors of finite efficiency (see Appendix 2). [0664] Let the set {|S _{i} } (where i=1, 2, . . . , N) be the search elements that a quantum computer apparatus is to deal with. Let these elements be indexed from 0 to N−1. This index can be stored in n bits where N≡2^{n}. Let the search problem have exactly M solutions with 1≦M≦N. Let f(ξ) be a function with ξ an integer in the range 0 to N−1. By definition, f(ξ)=1 if ξ is a solution to the search problem and f(ξ)=0 if ξ is not a solution to the search problem. One then needs an oracle that is able to recognize solutions to the search problem (see Appendix 3). This is signaled by making use of a qubit.
[0665] The oracle is a unitary operator 'ο defined by its action on computational basis as follows: 'ο:|ξ |q→|ξ|q⊕f(ξ)[0666] where |ξ is the index register and the oracle qubit |q is a single qubit that is flipped if f(ξ)=1 and is unchanged otherwise (see Appendix 4). Thus,[0667] |ξ |0→|ξ|0 if |ξ is not a solution[0668] |ξ |0→|ξ|1 if |ξ is a solution[0669] It is convenient to apply the oracle with the oracle qubit initially in the state
[0670] so that 'ο: |ξ |q→(−1)^{f(ξ)}|ξ|q. Then the oracle marks the solutions to the search by shifting the phase of the solution (see Appendices 3 and 4). If there are M solutions, it turns out that one need only apply the search oracle
[0671] times on the QC. Initially, the QC, assumed to be an integral part of the final detector, is always in the state |0 ^{ n}. The first step in the Grover's QSA is to apply a Hadamard transform to put the computer in the equal superposition state
[0672] The search algorithm then involves repeated applications of the Grover's iteration (or Grover's operator G) which can be broken up into the following four operations: 1) The oracle 'ο; 2) The Hadamard transform H ^{f(ξ)}|ξ; 4) The Hadamard transform H^{ n}.
[0673] The combined effect of steps 2, 3 and 4 is (see Appendix 3) I)H ^{ n}=2|ψ ψ|−I
[0674] where
[0675] The Grover's operator G can be regarded as a rotation in the two dimensional space spanned by the vector |ψ (see Appendices 3 and 4) which is a uniform superposition of the solutions to the search problem. To see this, define the normalized states[0676] where
[0677] indicates a sum over all ξ that are solutions to the search problem [0678] and
[0679] a sum over all, that are not solutions to the search problem. The initial state are not solutions can be written as
[0680] so that the apparatus (with quantum computing) is the space spanned by |α and |β to start with. Now notice (according to Appendix 4) that the oracle operator performs a rotation about the vector |α in the plane defined by |α and |β, i.e.,'ο( +b|β )=a|α−b|β.
[0681] Similarly, G also performs a reflection in the same plane about the vector |ψ , and the effect of these two reflections is a rotation. Therefore, the state Gk|ψ remains in the plane spanned by |α and |β for all k. The rotation angle can be found as follows. Let
[0682] so that
[0683] Then one can show (see Appendix 4) that
[0684] so that θ is indeed the rotation angle, and so
[0685] Thus, the repeated applications of the Grover's operator are rotated the vector |ψ close to |β.[0686] When this happens, an observation in the computational basis produces one of the outcomes superposed in |β with high probability. In a quantum measurement, only one outcome must occur and hence, the number M of simultaneous solutions that Grover's QSA searches is unity.[0687] 11.2. Grover's Search Algorithm and Quantum Lower Bounds. Searching an item in an unsorted DB with size N costs a classical computer O(N) running time. A search algorithm consults the DB only O(AN) times. In contrast to algorithms based on the quantum Fourier transformation, with exponential speed-up, the search algorithm only provides a quadratic improvement. However, the algorithm is important because it has broad applications and the same technique can be used to improve solutions of NP-complete problems. Grover's search algorithm is optimal. At least Ω({square root}{square root over (N)}) queries are needed to solve the problem. The following example illustrates the QSA and its lower bound respectively (see Appendices 4 and 5). [0688] Let f:[N]→{0, 1} be a Boolean function. Assume a quantum black box U _{f}:|x|0→|x|f(x).
[0689] If |y is initialized to[0690] the oracle acts as
[0691] Assume that there is a single value k such that f(k)=1. If f is specified by a black box, the lower bound is the fewest queries needed to f to determine k. [0692] 11.2.1. Inversion about the average and its application in the iterative procedure. The unitary transform
[0693] where E is the average of {a [0694] Appendix 4 describes the properties of the operator D. [0695] As shown in FIG. 67, the operator D increases (decreases) amplitudes that are originally below (above) the mean value μ. [0696] The QSA iteratively improves the probability of measuring a solution. In each iteration, this algorithm performs two operations: first it consults the oracle U |φ _{f}|φ_{i}
[0697] along with iteration i to iteration (i+1). [0698] For example, assume it is desired to find one out of N items. In the first step, as shown in FIG. 68A, prepare the initial state as a uniform superposition over these N items. In each iteration, the entanglement operator U [0699] For example, after the first iteration,
[0700] after the second iteration,
[0701] More formally, at iteration t, α [0702] Initially, α [0703] Increasing the number of iterations does not always increase the chance of measuring the right answer. The amplitude of the marked solution goes up and down as a cycle. If the iterations are not stopped at the right time, the chance of measuring the correct item is reduced. [0704] 11.2.2. The geometric interpretation. When finding M solutions from a sample space with N entries, one can cluster these items into two orthogonal bases, say
[0705] (the collection of the M solutions) and
[0706] (the collection of the remaining items). [0707]FIG. 69 helps in visualizing the iterative steps in a single plane spanned by these two vectors. [0708] For original state
[0709] according to Eq.(11.1), it can be rewritten as
[0710] In the oracle consultation, the operator U _{0} . The product of these two operators, DU_{f}, performs an equivalent 2θ-rotation operation, where
[0711] After i such iterations, the state becomes ( j+1)θ)|k +cos((2j+1)θ)|u
[0712] In the special case of N items (N>>1), θ≧sin θ=1/{square root}{square root over (N)}, to maximize the probability of obtaining the correct measurement, the needed number of iterations is:
[0713] Consequently, Grover's search algorithm makes O({square root}{square root over (N)}) queries. [0714] Through this visualization, it can be seen that if the number of iterations is not chosen properly, the final vector might not be rotated to a desired angle, which results in a small magnitude is projected onto the |k direction, which means a small probability of measuring the right answer.[0715] 11.2.3 Quantum Lower Bounds. In light of the previously-developed quantum algorithms, one might ask if a quantum computer can solve NP-complete problems in polynomial time. Consider the satisfiability (SAT) problem, the first proven NP-complete problem. It can be formulated as a search problem. That is, given a Boolean formula f(x [0716] For the hybrid argument, consider any quantum algorithm A for solving the search problem. First do a test run of A on function f≡0. Define the query magnitude of x to be
[0717] where α [0718] For such an x, by Cauchy-Schwarz inequality,
[0719] Let |φ _{1} , . . . , |φ_{T} be the states of A_{f}. Now run the algorithm A on the function g: g(x)=1, g(y)=0 ∀y≠x. Then ∥|φ_{T} −|ψ_{T} ∥ must be small.
[0720] It can be shown that |ψ _{T} +|E_{0} +|E_{1} + . . . +|E_{T−1} , where ∥|E_{t} ∥≦|α_{x,t}|. To show this, consider two runs of algorithm A, which differ only on the t-th step: one queries the function f and the other queries the function g. Both runs query the functionf in the first t−1 steps. Then at the end of the t-th step, the state of the first run is |φ_{t} , whereas the state of the second run is |φ_{t} +|F_{t} , where ∥|F_{t} ∥≦|α_{x,t}|. Now, if U is the unitary transform describing the remaining (T−t) steps, then the final state after T steps for the two runs are U|φ_{t} and U(|φ_{t} +|F_{t} ), respectively. The latter state can be written as U|φ_{t} +|E_{t} , where |E_{t} =U|F_{t} . Thus switching the queried function only on the t-th step results in a change in the final state of the algorithm by |F_{t} , where ∥|E_{t} ∥≦|α_{x,t}|. Therefore switching the queried function in all the steps results in the change |E_{0} +|E_{1} + . . . +|E_{T−}1 in the final state, where ∥|E_{t} ∥≦|α_{x,t}|.
[0721] It follows that
[0722] Measuring |ψ [0723] of the distribution that results from measuring |φ [0724] Thus, any algorithm that distinguishes f from g with constant probability must take a number of steps T=Ω({square root}{square root over (N)}). [0725] One can repeat the argument with another function h, and thus show that the final state of A while querying h satisfies
[0726] By the triangle inequality it is true that
[0727] Thus any quantum algorithm that distinguishes h from g with constant probability must take a number of steps T=Ω({square root}{square root over (N)}). [0728] For the quantum adversary method assume initially one has two unentangled registers, an input-register and a work-register, with states
[0729] And a quantum algorithm queries the first register and operates on the second register. If the algorithm works correctly, the final states of these registers must be strongly entangled, that is
[0730] The amplitudes in the above two expressions are omitted. A suitable measure of entanglement increases from 0 to N in queries from the initial state to final state. Moreover, the entanglement can only increase during a query and this increase is bounded by O({square root}{square root over (N)}) per query, thus yielding as Ω({square root}{square root over (N)}) a lower bound (see Appendices 3 and 4). [0731] 11.3. The forming of unified teaching signal by Grover's QSA. FIG. 21 shows the forming process of a KB of fuzzy P-controller in the ICSS. The box |